Spaces:

build-small-hackathon
/

Cook_with_a_LLM

Paused

App Files Files Community

FredinVázquez commited on Jun 12

Commit

bad5d84

1 Parent(s): c795fe5

add strategy plan

Browse files

Files changed (4) hide show

Strategy/arquitectura.html +668 -0
Strategy/estrategia.md +496 -0
Strategy/plan.md +245 -0
Strategy/plan_implementacion.md +674 -0

Strategy/arquitectura.html ADDED Viewed

	@@ -0,0 +1,668 @@

+<!DOCTYPE html>
+<html lang="es">
+<head>
+<meta charset="UTF-8" />
+<meta name="viewport" content="width=device-width, initial-scale=1.0" />
+<title>Cocina Conmigo — Plan visual del proyecto</title>
+<style>
+  :root {
+    --bg: #f5ecd9;
+    --card: #fffbf0;
+    --ink: #2b2018;
+    --accent: #a85c2a;       /* terracotta */
+    --accent-soft: #f6dccc;
+    --accent2: #6b4a2a;
+    --gold: #c9962b;
+    --green: #3f7a3a;
+    --green-soft: #dbe9d8;
+    --red: #b03a2e;
+    --red-soft: #f4d6d2;
+    --gray: #8a7e6f;
+    --line: #d8c9ad;
+  }
+  * { box-sizing: border-box; }
+  body {
+    font-family: 'Inter', -apple-system, sans-serif;
+    background: var(--bg);
+    color: var(--ink);
+    margin: 0;
+    padding: 32px 16px 80px;
+    line-height: 1.55;
+  }
+  .wrap { max-width: 1240px; margin: 0 auto; }
+  h1 { font-family: 'Lora', Georgia, serif; font-size: 46px; margin: 0 0 4px;
+       letter-spacing: -0.5px; font-weight: 700; }
+  h1 em { color: var(--accent); font-style: italic; }
+  .subtitle { color: var(--accent2); font-style: italic; margin-bottom: 28px; font-size: 17px; }
+  h2 {
+    margin-top: 56px; border-top: 1px dashed var(--line); padding-top: 24px;
+    font-size: 26px; font-family: 'Lora', Georgia, serif; letter-spacing: 0.3px;
+  }
+  h2 .num {
+    color: var(--accent); font-family: ui-monospace, monospace;
+    font-size: 20px; margin-right: 10px;
+  }
+  h3 { font-size: 18px; margin-top: 28px; color: var(--accent2); font-family: 'Lora', Georgia, serif; }
+  /* Hero */
+  .hero {
+    background: var(--card); border: 2px solid var(--ink); border-radius: 14px;
+    padding: 30px 32px; display: grid; grid-template-columns: 1fr; gap: 18px;
+  }
+  @media(min-width: 760px){ .hero { grid-template-columns: 2fr 1fr; align-items: center; } }
+  .hero h2 { border:0; margin:0 0 6px; padding:0; font-size: 22px; }
+  .hero .quote {
+    font-style: italic; font-size: 17px; color: var(--accent2);
+    border-left: 3px solid var(--accent); padding-left: 14px; margin: 6px 0 0;
+  }
+  .hero .target {
+    background: #fff3cf; border-radius: 12px; padding: 14px 16px;
+    font-size: 13px; border: 1px solid var(--line); line-height: 1.55;
+  }
+  .hero .target strong { color: var(--accent); }
+  /* Pills */
+  .pill {
+    display: inline-block; padding: 2px 9px; border-radius: 12px;
+    color: white; font-size: 12px; margin: 2px 4px 2px 0; font-family: ui-monospace, monospace;
+  }
+  .pill.user    { background: var(--gray); }
+  .pill.gradio  { background: var(--accent); }
+  .pill.hf      { background: var(--gold); }
+  .pill.modal   { background: var(--green); }
+  .pill.flux    { background: #111; }
+  .pill.openbmb { background: #075e54; }
+  .pill.cohere  { background: #5e3aa3; }
+  .pill.openai  { background: #2c5e8a; }
+  .pill.llama   { background: #6a3d8a; }
+  /* Phone/recipe card mockup */
+  .phone-row {
+    display: grid; grid-template-columns: 1fr; gap: 18px; margin-top: 16px;
+  }
+  @media(min-width: 760px){ .phone-row { grid-template-columns: repeat(4, 1fr); } }
+  .phone {
+    background: #111; border-radius: 24px; padding: 8px;
+    box-shadow: 0 8px 22px rgba(0,0,0,0.18);
+  }
+  .phone .screen {
+    background: #fffbf0; border-radius: 18px; overflow: hidden;
+    height: 380px; display: flex; flex-direction: column;
+  }
+  .phone .topbar {
+    background: var(--accent); color: white; padding: 10px 14px;
+    font-size: 13px; font-family: 'Lora', serif;
+  }
+  .phone .body { padding: 12px; flex: 1; overflow-y: auto; font-size: 12px; }
+  .phone .body .illu {
+    width: 100%; aspect-ratio: 4/3; border-radius: 8px;
+    background: linear-gradient(135deg, #ffd28b 0%, #c97a3e 100%);
+    display: flex; align-items: center; justify-content: center;
+    font-size: 48px; box-shadow: 0 2px 8px rgba(0,0,0,0.1); margin-bottom: 8px;
+  }
+  .phone .body p { margin: 6px 0; line-height: 1.5; }
+  .phone .body .voice {
+    background: var(--green-soft); border-radius: 6px; padding: 6px 10px;
+    margin-top: 8px; font-size: 11px; color: var(--green);
+  }
+  .phone .body .tip {
+    background: var(--red-soft); border-radius: 6px; padding: 6px 10px;
+    margin-top: 6px; font-size: 11px; color: var(--red);
+  }
+  .scenario-label {
+    text-align: center; font-size: 13px; color: var(--accent2);
+    margin-top: 8px; font-style: italic;
+  }
+  /* SVG */
+  svg { width: 100%; height: auto; display: block; }
+  .node-box { fill: var(--card); stroke: var(--ink); stroke-width: 1.5; }
+  .node-text { font-family: 'Inter', sans-serif; font-size: 14px; fill: var(--ink); }
+  .node-title { font-weight: 700; font-size: 15px; }
+  .node-sub { font-size: 11px; fill: var(--accent2); font-style: italic; }
+  .arrow { stroke: var(--ink); stroke-width: 1.8; fill: none; }
+  .arrow-label { font-size: 11px; fill: var(--accent2); font-family: ui-monospace, monospace; }
+  .dashed { stroke-dasharray: 6 4; }
+  .arrow-loop { stroke: var(--accent); stroke-width: 2.2; fill: none; }
+  /* Cards */
+  .grid-2 { display: grid; grid-template-columns: 1fr; gap: 18px; margin-top: 16px; }
+  @media(min-width: 880px){ .grid-2 { grid-template-columns: 1fr 1fr; } }
+  .grid-3 { display: grid; grid-template-columns: 1fr; gap: 14px; margin-top: 14px; }
+  @media(min-width: 760px){ .grid-3 { grid-template-columns: repeat(3, 1fr); } }
+  .card {
+    background: var(--card); border: 1px solid var(--line);
+    border-radius: 10px; padding: 18px 20px;
+  }
+  .card.pick { border: 2px solid var(--accent); }
+  .pick-tag {
+    display: inline-block; background: var(--accent); color: white;
+    font-family: ui-monospace, monospace; font-size: 11px;
+    padding: 1px 7px; border-radius: 10px; margin-bottom: 6px;
+  }
+  table {
+    width: 100%; border-collapse: collapse; background: var(--card);
+    border: 1px solid var(--line); margin-top: 14px; font-size: 14px;
+  }
+  th, td { padding: 8px 10px; text-align: left; border-bottom: 1px solid var(--line); vertical-align: top; }
+  th { background: #efe4cb; font-size: 13px; letter-spacing: 0.5px; text-transform: uppercase; }
+  code {
+    background: #efe4cb; border-radius: 3px; padding: 1px 5px; font-size: 13px;
+  }
+  /* Forbidden zone */
+  .forbidden {
+    background: var(--red-soft); border: 1px solid var(--red);
+    border-radius: 8px; padding: 14px 18px; margin-top: 14px;
+  }
+  .forbidden strong { color: var(--red); }
+  .forbidden ul {
+    columns: 2; column-gap: 28px; margin: 8px 0 0; padding-left: 18px; font-size: 14px;
+  }
+  /* Timeline */
+  .timeline { position: relative; padding-left: 36px; margin-top: 20px; }
+  .timeline::before {
+    content: ""; position: absolute; left: 12px; top: 6px; bottom: 6px;
+    width: 3px; background: var(--accent); border-radius: 2px;
+  }
+  .day {
+    position: relative; margin-bottom: 14px; background: var(--card);
+    border: 1px solid var(--line); border-radius: 8px; padding: 12px 16px;
+  }
+  .day::before {
+    content: ""; position: absolute; left: -29px; top: 16px;
+    width: 13px; height: 13px; background: var(--accent);
+    border: 2px solid var(--card); border-radius: 50%;
+  }
+  .day .lbl {
+    display: inline-block; background: var(--accent); color: white;
+    font-family: ui-monospace, monospace; font-size: 11px;
+    padding: 1px 7px; border-radius: 10px; margin-right: 8px;
+  }
+  .day strong { font-size: 15px; }
+  .day .what { font-size: 13px; color: var(--accent2); margin-top: 2px; }
+  /* Award rows */
+  .award-row {
+    display: flex; justify-content: space-between;
+    padding: 8px 12px; border-bottom: 1px solid var(--line); font-size: 14px;
+  }
+  .award-row:last-child { border-bottom: 0; }
+  .prob {
+    font-family: ui-monospace, monospace; font-size: 12px;
+    padding: 1px 8px; border-radius: 10px; color: white;
+  }
+  .prob-h { background: #2e7d32; }
+  .prob-m { background: #ef9c2c; }
+  .prob-l { background: #b03a2e; }
+  /* Badges grid */
+  .badges-grid {
+    display: grid; grid-template-columns: repeat(auto-fit, minmax(180px, 1fr));
+    gap: 12px; margin-top: 14px;
+  }
+  .badge-card {
+    background: var(--card); border: 1px solid var(--line);
+    border-radius: 8px; padding: 12px 14px;
+  }
+  .badge-card.skip { opacity: 0.45; border-style: dashed; }
+  .badge-card .tag {
+    display: inline-block; background: var(--accent); color: white;
+    font-family: ui-monospace, monospace; font-size: 11px;
+    padding: 1px 7px; border-radius: 10px; margin-bottom: 6px;
+  }
+  .badge-card.skip .tag { background: var(--gray); }
+  .badge-card strong { font-size: 14px; }
+  .badge-card p { font-size: 13px; color: var(--accent2); margin: 4px 0 0; }
+  .footnote {
+    margin-top: 30px; padding: 14px 18px;
+    border-left: 4px solid var(--accent);
+    background: var(--card); font-size: 14px; border-radius: 4px;
+  }
+</style>
+<link rel="preconnect" href="https://fonts.googleapis.com">
+<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
+<link href="https://fonts.googleapis.com/css2?family=Lora:wght@400;600;700&family=Inter:wght@400;500;600;700&display=swap" rel="stylesheet">
+</head>
+<body>
+<div class="wrap">
+  <h1><em>Cocina Conmigo</em></h1>
+  <div class="subtitle">Sous-chef multimodal con visión, voz y Flux.2 — para cocinar con tu mamá sin tener las manos libres</div>
+  <div class="hero">
+    <div>
+      <h2>La idea en una frase</h2>
+      <p>Tu mamá toma foto del refri, la app le propone qué cocinar, le <strong>muestra cómo se debe ver cada paso</strong> con Flux.2, y la <strong>narra por voz</strong> mientras ella cocina con las manos llenas.</p>
+      <p class="quote">"Mi mamá me pidió que le enseñara a hacer ramen. Le construí un sous-chef que vive en su tablet."</p>
+      <div style="margin-top: 14px;">
+        <span class="pill flux">Flux.2 Klein 9B</span>
+        <span class="pill openbmb">MiniCPM-V + voice</span>
+        <span class="pill cohere">Cohere voice</span>
+        <span class="pill gradio">Gradio Workflows</span>
+        <span class="pill modal">Modal-powered</span>
+        <span class="pill llama">llama.cpp</span>
+      </div>
+    </div>
+    <div class="target">
+      <strong>Track:</strong> Backyard AI<br/>
+      <strong>Persona:</strong> tu mamá / pareja / vecino<br/>
+      <strong>Idioma:</strong> español-mexicano<br/>
+      <strong>Total params:</strong> ~17B (≤ 32B ✓)<br/>
+      <strong>Cocina:</strong> mexicana tradicional<br/>
+      <strong>Storyline:</strong> "Para que mi mamá deje de googlear"
+    </div>
+  </div>
+  <h2><span class="num">01</span>Por qué esta idea, y no las anteriores</h2>
+  <table>
+    <thead><tr><th>Iteración</th><th>Idea</th><th>Por qué se descartó</th></tr></thead>
+    <tbody>
+      <tr><td>v1</td><td>Abuelita (parent phone helper)</td><td>En la lista pre-cocinada de OpenBMB → 5-15 equipos lo harán</td></tr>
+      <tr><td>v2</td><td>Cuentacuentos (voice storyteller)</td><td>También en la lista pre-cocinada de OpenBMB</td></tr>
+      <tr style="background:#fff3cf;"><td><strong>v3 (ésta)</strong></td><td><strong>Cocina Conmigo</strong></td><td>Refinamiento de tu idea #1 · NO está en ninguna lista pre-cocinada · usa Flux.2 + Workflows + voces · diaria + universal</td></tr>
+    </tbody>
+  </table>
+  <div class="forbidden">
+    <strong>⛔ Las 12 ideas en zona prohibida (clúster OpenBMB):</strong>
+    <ul>
+      <li>parent phone helper</li>
+      <li>receipt / bill explainer</li>
+      <li>shop menu / repair manual</li>
+      <li>offline personal assistant / voice companion</li>
+      <li>voice storyteller</li>
+      <li>visual mystery box</li>
+      <li>AI museum (≈ tu idea #4)</li>
+      <li>doodle creature</li>
+      <li>dream postcard gen</li>
+      <li>omni-modal adventure</li>
+      <li>tiny local NPC / character agent</li>
+      <li>cortes de cabello (tu idea #3, ya saturada)</li>
+    </ul>
+  </div>
+  <h2><span class="num">02</span>Las 4 historias del demo</h2>
+  <div class="phone-row">
+    <div>
+      <div class="phone"><div class="screen">
+        <div class="topbar">📸 Tengo esto en el refri</div>
+        <div class="body">
+          <div class="illu">🍅🌶🐔🧅</div>
+          <p><strong>Veo:</strong> pollo, jitomate, cebolla, cilantro, tortillas, queso.</p>
+          <p style="background:#fff3cf;border-radius:6px;padding:6px 10px;">
+            <strong>3 opciones:</strong><br/>
+            🌮 Tinga · 🌯 Enchiladas · 🧀 Quesadillas
+          </p>
+          <div class="voice">🔊 "¿Qué traes ganas?"</div>
+        </div>
+      </div></div>
+      <div class="scenario-label">1. Visión + Planner</div>
+    </div>
+    <div>
+      <div class="phone"><div class="screen">
+        <div class="topbar">👩‍🍳 Paso 2 de 5</div>
+        <div class="body">
+          <div class="illu">🍳✨</div>
+          <p><strong>Acitrona la cebolla en aceite caliente.</strong></p>
+          <p style="font-size:11px;color:var(--gray);">⏱ 4 minutos · hasta que esté transparente</p>
+          <div class="voice">🔊 OpenBMB voice narra…</div>
+        </div>
+      </div></div>
+      <div class="scenario-label">2. Voz + imagen objetivo</div>
+    </div>
+    <div>
+      <div class="phone"><div class="screen">
+        <div class="topbar">📸 ¿Voy bien?</div>
+        <div class="body">
+          <div class="illu">🍳👀</div>
+          <p style="background:var(--green-soft);border-radius:6px;padding:6px 10px;color:var(--green);">
+            <strong>✅ Va perfecto.</strong> La cebolla ya se ve transparente.
+          </p>
+          <div class="tip">🔊 Cohere voice: "¡Súbele 1 minuto más, está bien!"</div>
+        </div>
+      </div></div>
+      <div class="scenario-label">3. Closed-loop visual</div>
+    </div>
+    <div>
+      <div class="phone"><div class="screen">
+        <div class="topbar">🔄 Replan</div>
+        <div class="body">
+          <p>Usuario: <em>"No tengo cilantro."</em></p>
+          <div class="illu" style="background: linear-gradient(135deg,#ffd28b,#a85c2a);">🌮</div>
+          <p>"No pasa nada. Le ponemos perejil o nada. Sigue siendo tinga."</p>
+          <div class="voice">🔊 Receta regenera · plato final actualizado</div>
+        </div>
+      </div></div>
+      <div class="scenario-label">4. Adaptación en vivo</div>
+    </div>
+  </div>
+  <h2><span class="num">03</span>Arquitectura — 5 agentes en un Gradio Workflow</h2>
+  <svg viewBox="0 0 1240 540" xmlns="http://www.w3.org/2000/svg">
+    <defs>
+      <marker id="ar" viewBox="0 0 10 10" refX="9" refY="5" markerWidth="7" markerHeight="7" orient="auto">
+        <path d="M0,0 L10,5 L0,10 z" fill="#2b2018"/>
+      </marker>
+      <marker id="aro" viewBox="0 0 10 10" refX="9" refY="5" markerWidth="7" markerHeight="7" orient="auto">
+        <path d="M0,0 L10,5 L0,10 z" fill="#a85c2a"/>
+      </marker>
+    </defs>
+    <!-- User input area -->
+    <rect x="20" y="40" width="200" height="240" rx="10" fill="#fff3cf" stroke="#d8c9ad" stroke-dasharray="4 3"/>
+    <text x="40" y="62" class="node-text node-title" fill="#6b4a2a">USUARIO (cocina)</text>
+    <rect class="node-box" x="40" y="80" width="160" height="50" rx="6" fill="#ddd1bd"/>
+    <text x="120" y="102" class="node-text node-title" text-anchor="middle">📸 Foto del refri</text>
+    <text x="120" y="118" class="node-text node-sub" text-anchor="middle">trigger inicial</text>
+    <rect class="node-box" x="40" y="140" width="160" height="50" rx="6" fill="#ddd1bd"/>
+    <text x="120" y="162" class="node-text node-title" text-anchor="middle">🎙️ Pregunta voz</text>
+    <text x="120" y="178" class="node-text node-sub" text-anchor="middle">"¿voy bien?"</text>
+    <rect class="node-box" x="40" y="200" width="160" height="50" rx="6" fill="#ddd1bd"/>
+    <text x="120" y="222" class="node-text node-title" text-anchor="middle">📸 Foto progreso</text>
+    <text x="120" y="238" class="node-text node-sub" text-anchor="middle">closed-loop</text>
+    <!-- Output area -->
+    <rect x="20" y="320" width="200" height="180" rx="10" fill="#fff3cf" stroke="#d8c9ad" stroke-dasharray="4 3"/>
+    <text x="40" y="342" class="node-text node-title" fill="#6b4a2a">SALIDA</text>
+    <rect class="node-box" x="40" y="360" width="160" height="50" rx="6" fill="#dbe9d8"/>
+    <text x="120" y="382" class="node-text node-title" text-anchor="middle">🍽️ Plato final + receta</text>
+    <text x="120" y="398" class="node-text node-sub" text-anchor="middle">imagen + texto</text>
+    <rect class="node-box" x="40" y="420" width="160" height="50" rx="6" fill="#dbe9d8"/>
+    <text x="120" y="442" class="node-text node-title" text-anchor="middle">🔊 Voz por paso</text>
+    <text x="120" y="458" class="node-text node-sub" text-anchor="middle">narrador + tips</text>
+    <!-- Pipeline center -->
+    <rect x="260" y="40" width="700" height="460" rx="10" fill="#fffaf0" stroke="#d8c9ad" stroke-width="1.5"/>
+    <text x="610" y="62" class="node-text node-title" text-anchor="middle" fill="#6b4a2a">HF SPACE — Gradio Workflow (5 agentes)</text>
+    <!-- Vision (Mise en Place) -->
+    <rect class="node-box" x="280" y="90" width="200" height="80" rx="6" fill="#e6d5ed"/>
+    <text x="380" y="110" class="node-text node-title" text-anchor="middle">👁️ MISE EN PLACE</text>
+    <text x="380" y="126" class="node-text node-sub" text-anchor="middle">MiniCPM-V (Q4)</text>
+    <text x="380" y="142" class="node-text node-sub" text-anchor="middle">~2-4B</text>
+    <text x="380" y="160" class="node-text node-sub" text-anchor="middle">identifica ingredientes</text>
+    <!-- Recipe Planner -->
+    <rect class="node-box" x="510" y="90" width="200" height="80" rx="6" fill="#fbe4d3"/>
+    <text x="610" y="110" class="node-text node-title" text-anchor="middle">🧠 RECIPE PLANNER</text>
+    <text x="610" y="126" class="node-text node-sub" text-anchor="middle">MiniCPM-4 (LoRA mx)</text>
+    <text x="610" y="142" class="node-text node-sub" text-anchor="middle">~4B</text>
+    <text x="610" y="160" class="node-text node-sub" text-anchor="middle">arma receta JSON · replan</text>
+    <!-- Step Illustrator -->
+    <rect class="node-box" x="740" y="90" width="200" height="80" rx="6" fill="#f6dccc"/>
+    <text x="840" y="110" class="node-text node-title" text-anchor="middle">🎨 STEP ILLUSTRATOR</text>
+    <text x="840" y="126" class="node-text node-sub" text-anchor="middle">Flux.2 Klein 9B</text>
+    <text x="840" y="142" class="node-text node-sub" text-anchor="middle">en Modal GPU L4</text>
+    <text x="840" y="160" class="node-text node-sub" text-anchor="middle">imagen-objetivo por paso</text>
+    <!-- Sous-Chef Narrator -->
+    <rect class="node-box" x="510" y="200" width="200" height="70" rx="6" fill="#cfe0ee"/>
+    <text x="610" y="222" class="node-text node-title" text-anchor="middle">🔊 SOUS-CHEF NARRATOR</text>
+    <text x="610" y="238" class="node-text node-sub" text-anchor="middle">OpenBMB voice (~1B)</text>
+    <text x="610" y="254" class="node-text node-sub" text-anchor="middle">tono cálido</text>
+    <!-- Tip Giver -->
+    <rect class="node-box" x="740" y="200" width="200" height="70" rx="6" fill="#e9d6f5"/>
+    <text x="840" y="222" class="node-text node-title" text-anchor="middle">🎭 TIP GIVER</text>
+    <text x="840" y="238" class="node-text node-sub" text-anchor="middle">Cohere voice (~1B)</text>
+    <text x="840" y="254" class="node-text node-sub" text-anchor="middle">warnings · enérgico</text>
+    <!-- Progress Validator (closed loop) -->
+    <rect class="node-box" x="280" y="290" width="220" height="90" rx="6" fill="#dbe9d8" stroke="#3f7a3a" stroke-width="2"/>
+    <text x="390" y="312" class="node-text node-title" text-anchor="middle" fill="#3f7a3a">✅ PROGRESS VALIDATOR</text>
+    <text x="390" y="328" class="node-text node-sub" text-anchor="middle">MiniCPM-V (reuso)</text>
+    <text x="390" y="344" class="node-text node-sub" text-anchor="middle">compara foto usuario vs</text>
+    <text x="390" y="360" class="node-text node-sub" text-anchor="middle">imagen-objetivo</text>
+    <text x="390" y="376" class="node-text node-sub" text-anchor="middle">CLOSED LOOP 🔄</text>
+    <!-- STT -->
+    <rect class="node-box" x="280" y="200" width="200" height="70" rx="6" fill="#cfe0ee"/>
+    <text x="380" y="222" class="node-text node-title" text-anchor="middle">🎙️ STT (opcional)</text>
+    <text x="380" y="238" class="node-text node-sub" text-anchor="middle">Whisper-tiny (~40M)</text>
+    <text x="380" y="254" class="node-text node-sub" text-anchor="middle">"¿voy bien?" hands-free</text>
+    <!-- Recipe State -->
+    <rect class="node-box" x="510" y="290" width="430" height="90" rx="6" fill="#fff3cf"/>
+    <text x="725" y="312" class="node-text node-title" text-anchor="middle" fill="#8a6a18">📖 RECIPE STATE (dataclass)</text>
+    <text x="725" y="328" class="node-text node-sub" text-anchor="middle">name · final_dish_image · steps · current_step ·</text>
+    <text x="725" y="344" class="node-text node-sub" text-anchor="middle">missing_ingredients · substitutes · user_progress_photos</text>
+    <text x="725" y="362" class="node-text node-sub" text-anchor="middle">cada agente lee y escribe sobre este objeto</text>
+    <!-- Page assembler -->
+    <rect class="node-box" x="280" y="400" width="660" height="60" rx="6" fill="#f6dccc"/>
+    <text x="610" y="422" class="node-text node-title" text-anchor="middle">📖 RECIPE CARD ASSEMBLER</text>
+    <text x="610" y="438" class="node-text node-sub" text-anchor="middle">renderiza la tarjeta de receta + cards por paso + audio reproducible</text>
+    <!-- Modal box -->
+    <rect x="990" y="40" width="240" height="460" rx="10" fill="#dbe9d8" stroke="#3f7a3a" stroke-width="1.5"/>
+    <text x="1110" y="62" class="node-text node-title" text-anchor="middle" fill="#3f7a3a">MODAL</text>
+    <rect class="node-box" x="1010" y="90" width="200" height="80" rx="6" fill="#fff"/>
+    <text x="1110" y="112" class="node-text node-title" text-anchor="middle">Flux endpoint</text>
+    <text x="1110" y="128" class="node-text node-sub" text-anchor="middle">runtime · @app.cls L4</text>
+    <text x="1110" y="144" class="node-text node-sub" text-anchor="middle">scaledown 180s</text>
+    <text x="1110" y="160" class="node-text node-sub" text-anchor="middle">~1-3s/imagen</text>
+    <rect class="node-box" x="1010" y="190" width="200" height="80" rx="6" fill="#fff"/>
+    <text x="1110" y="212" class="node-text node-title" text-anchor="middle">Dataset cocina mx</text>
+    <text x="1110" y="228" class="node-text node-sub" text-anchor="middle">offline · 200 recetas</text>
+    <text x="1110" y="244" class="node-text node-sub" text-anchor="middle">Codex API genera</text>
+    <text x="1110" y="260" class="node-text node-sub" text-anchor="middle">~$5</text>
+    <rect class="node-box" x="1010" y="290" width="200" height="80" rx="6" fill="#fff"/>
+    <text x="1110" y="312" class="node-text node-title" text-anchor="middle">LoRA Planner</text>
+    <text x="1110" y="328" class="node-text node-sub" text-anchor="middle">offline · A10G ~30 min</text>
+    <text x="1110" y="344" class="node-text node-sub" text-anchor="middle">push GGUF a HF</text>
+    <text x="1110" y="360" class="node-text node-sub" text-anchor="middle">~$1</text>
+    <rect class="node-box" x="1010" y="390" width="200" height="80" rx="6" fill="#fff"/>
+    <text x="1110" y="412" class="node-text node-title" text-anchor="middle">Eval pipeline</text>
+    <text x="1110" y="428" class="node-text node-sub" text-anchor="middle">consistencia visual</text>
+    <text x="1110" y="444" class="node-text node-sub" text-anchor="middle">% ingredientes correctos</text>
+    <text x="1110" y="460" class="node-text node-sub" text-anchor="middle">~$1</text>
+    <!-- Arrows: input → vision -->
+    <path class="arrow" d="M200 105 L278 130" marker-end="url(#ar)"/>
+    <text x="200" y="100" class="arrow-label">refri</text>
+    <!-- input → STT -->
+    <path class="arrow" d="M200 165 L278 235" marker-end="url(#ar)"/>
+    <text x="205" y="200" class="arrow-label">audio</text>
+    <!-- input progress → validator -->
+    <path class="arrow arrow-loop" d="M200 225 L278 330" marker-end="url(#aro)"/>
+    <text x="200" y="270" class="arrow-label" style="fill:#a85c2a;">progreso</text>
+    <!-- Vision → Planner -->
+    <path class="arrow" d="M480 130 L508 130" marker-end="url(#ar)"/>
+    <text x="482" y="120" class="arrow-label">ingredientes</text>
+    <!-- Planner → Illustrator -->
+    <path class="arrow" d="M710 130 L738 130" marker-end="url(#ar)"/>
+    <text x="712" y="120" class="arrow-label">visual prompt</text>
+    <!-- Illustrator → Modal -->
+    <path class="arrow dashed" d="M940 130 L1008 130" marker-end="url(#ar)"/>
+    <text x="945" y="120" class="arrow-label">.remote()</text>
+    <!-- Planner → narrator -->
+    <path class="arrow" d="M610 170 L610 198" marker-end="url(#ar)"/>
+    <!-- Planner → tip giver -->
+    <path class="arrow" d="M710 145 C 760 170, 800 180, 800 198" marker-end="url(#ar)"/>
+    <!-- Validator → Planner (loop) -->
+    <path class="arrow arrow-loop" d="M390 290 C 390 240, 470 190, 510 145" marker-end="url(#aro)"/>
+    <text x="395" y="240" class="arrow-label" style="fill:#a85c2a;">verdict · feedback</text>
+    <!-- STT → Validator -->
+    <path class="arrow dashed" d="M380 270 L380 288" marker-end="url(#ar)"/>
+    <!-- Recipe state ↔ all agents -->
+    <path class="arrow dashed" d="M725 290 L725 270" marker-end="url(#ar)"/>
+    <path class="arrow dashed" d="M610 290 L610 270" marker-end="url(#ar)"/>
+    <!-- All → Assembler -->
+    <path class="arrow" d="M610 380 L610 398" marker-end="url(#ar)"/>
+    <!-- Assembler → output -->
+    <path class="arrow" d="M280 425 C 240 425, 220 410, 200 385" marker-end="url(#ar)"/>
+    <path class="arrow" d="M280 440 C 240 440, 220 445, 200 445" marker-end="url(#ar)"/>
+    <!-- Modal → Planner (LoRA pesos offline) -->
+    <path class="arrow dashed" d="M1010 330 C 870 330, 750 280, 710 165" marker-end="url(#ar)"/>
+    <text x="900" y="280" class="arrow-label">LoRA pesos</text>
+  </svg>
+  <p style="font-size: 13px; color: var(--accent2); margin-top: 10px;">
+    <strong>Flecha naranja</strong> = closed-loop visual (la innovación). El usuario toma foto del progreso, MiniCPM-V valida vs imagen-objetivo, el Planner ajusta o avanza. Ningún recipe app del mercado lo hace.
+  </p>
+  <h2><span class="num">04</span>El truco innovador: closed-loop visual cocinero</h2>
+  <div class="grid-3">
+    <div class="card">
+      <h3>1. Imagen-objetivo por paso</h3>
+      <p style="font-size:13px;">Flux.2 genera "así debe verse el sartén/plato/olla en el paso N". No es texto, no es stock photo: es generación context-aware del estado deseado.</p>
+    </div>
+    <div class="card">
+      <h3>2. Validación con foto del usuario</h3>
+      <p style="font-size:13px;">El usuario sube foto de cómo va. MiniCPM-V compara contra la imagen-objetivo y devuelve verdict: <code>go</code> · <code>wait</code> · <code>fix</code>.</p>
+    </div>
+    <div class="card">
+      <h3>3. Replan adaptativo</h3>
+      <p style="font-size:13px;">"No tengo cilantro." → Planner regenera receta + Flux regenera imagen final. El plan no es estático, evoluciona con el estado real.</p>
+    </div>
+  </div>
+  <p style="margin-top:14px; font-size:14px;">
+    <strong>Esta es la sección destacada del README</strong> y el blog post de Field Notes badge: <em>"How visual closed-loop cooking guidance works."</em>
+  </p>
+  <h2><span class="num">05</span>Badges objetivo (5/6)</h2>
+  <div class="badges-grid">
+    <div class="badge-card"><span class="tag">LLAMA.CPP</span><br/><strong>Llama Champion</strong><p>Vision + Planner via <code>llama-cpp-python</code> con GGUF Q4.</p></div>
+    <div class="badge-card"><span class="tag">FINE-TUNED</span><br/><strong>Well-Tuned</strong><p>LoRA en cocina mexicana · publicado en HF.</p></div>
+    <div class="badge-card"><span class="tag">CUSTOM UI</span><br/><strong>Off-Brand</strong><p>UI tarjeta de receta · serif · paleta cálida · modo cocina XL.</p></div>
+    <div class="badge-card"><span class="tag">OPEN TRACE</span><br/><strong>Sharing is Caring</strong><p>Dataset 150 recetas mx + traces + recetas generadas al Hub.</p></div>
+    <div class="badge-card"><span class="tag">TENTATIVE</span><br/><strong>Field Notes</strong><p>Blog: "Le construí un sous-chef a mi mamá".</p></div>
+    <div class="badge-card skip"><span class="tag">LOCAL-FIRST</span><br/><strong>Off the Grid</strong><p>Sacrificado: Flux.2 corre en Modal por calidad.</p></div>
+  </div>
+  <h2><span class="num">06</span>Premios objetivo</h2>
+  <div class="card">
+    <div class="award-row"><span><strong>Backyard AI Track</strong> · $1K–$4K</span><span class="prob prob-h">ALTA</span></div>
+    <div class="award-row"><span><strong>Modal Awards</strong> · $3K–$10K credits</span><span class="prob prob-h">ALTA</span></div>
+    <div class="award-row"><span><strong>OpenBMB Award</strong> · $1K–$2.5K</span><span class="prob prob-h">ALTA</span></div>
+    <div class="award-row"><span><strong>Best Demo</strong> · $1K</span><span class="prob prob-h">ALTA</span></div>
+    <div class="award-row"><span><strong>Community Choice</strong> · $2K</span><span class="prob prob-h">ALTA</span></div>
+    <div class="award-row"><span><strong>Best Agent</strong> · $1K</span><span class="prob prob-h">ALTA — closed-loop multi-agente real</span></div>
+    <div class="award-row"><span><strong>Bonus Quest Champion</strong> · $2K</span><span class="prob prob-m">MEDIA-ALTA · 5/6 badges</span></div>
+    <div class="award-row"><span><strong>Off-Brand</strong> · $1.5K</span><span class="prob prob-m">MEDIA</span></div>
+    <div class="award-row"><span><strong>Tiny Titan</strong> · $1.5K</span><span class="prob prob-l">BAJA · Flux 9B saca del rango</span></div>
+  </div>
+  <p style="font-size: 14px; margin-top: 8px;"><strong>Cota razonable acumulada: $5K–$12K cash + $3K–$10K Modal credits.</strong></p>
+  <h2><span class="num">07</span>Timeline de 10 días</h2>
+  <div class="timeline">
+    <div class="day"><span class="lbl">D1</span><strong>Setup + Modal Flux endpoint</strong><div class="what">"Hola Flux": prompt → imagen de un platillo. Space vacío deployado.</div></div>
+    <div class="day"><span class="lbl">D2</span><strong>Vision: identificación de ingredientes</strong><div class="what">MiniCPM-V Q4 · prueba con 5 fotos reales del refri.</div></div>
+    <div class="day"><span class="lbl">D3</span><strong>Recipe Planner LLM</strong><div class="what">MiniCPM-4 · JSON estructurado · 3 opciones a partir de ingredientes.</div></div>
+    <div class="day"><span class="lbl">D4</span><strong>Step Illustrator (Flux + consistencia)</strong><div class="what">Imagen del plato final + 5 imágenes-objetivo por paso · i2i suave.</div></div>
+    <div class="day"><span class="lbl">D5</span><strong>Voz: narrador + tip-giver</strong><div class="what">OpenBMB voice + Cohere voice · audio pre-renderizado por paso.</div></div>
+    <div class="day"><span class="lbl">D6</span><strong>UI Off-Brand: recipe card</strong><div class="what">gr.Blocks + CSS serif tierra · modo cocina XL hands-free.</div></div>
+    <div class="day"><span class="lbl">D7</span><strong>Gradio Workflows showcase</strong><div class="what">Pipeline reescrita como Workflow visible · pestaña separada.</div></div>
+    <div class="day"><span class="lbl">D8</span><strong>Fine-tune del Planner en cocina mx</strong><div class="what">200 recetas sintéticas · LoRA · GGUF · push HF.</div></div>
+    <div class="day"><span class="lbl">D9</span><strong>STT + Progress Validator + eval</strong><div class="what">Whisper · closed-loop activo · Sharing is Caring badge.</div></div>
+    <div class="day"><span class="lbl">D10</span><strong>Demo + README + blog + submit</strong><div class="what">Mamá real cocinando · 60-90s · subtítulos EN · Field Notes blog.</div></div>
+  </div>
+  <h2><span class="num">08</span>Plan B (corte de scope)</h2>
+  <table>
+    <thead><tr><th>#</th><th>Cortar</th><th>Pierdes</th><th>Conservas</th></tr></thead>
+    <tbody>
+      <tr><td>1</td><td>STT (preguntas voz)</td><td>comodidad demo</td><td>texto + foto</td></tr>
+      <tr><td>2</td><td>2da voz (Cohere tip-giver)</td><td>1 sponsor voice</td><td>narrador único</td></tr>
+      <tr><td>3</td><td>Progress Validator (closed-loop)</td><td><strong>Best Agent</strong> + innovación principal</td><td>demo lineal</td></tr>
+      <tr><td>4</td><td>Fine-tune del Planner</td><td><strong>Well-Tuned</strong></td><td>resto badges</td></tr>
+      <tr><td>5</td><td>Gradio Workflows showcase</td><td>diferenciador "fresh"</td><td>pipeline Python</td></tr>
+      <tr><td>6</td><td>UI super-custom</td><td><strong>Off-Brand</strong></td><td>UI default</td></tr>
+      <tr style="background:#fff3cf;"><td>—</td><td><strong>NUNCA</strong></td><td colspan="2">Vision + Planner + Illustrator + Narrator + UI mínima + video con persona real cocinando</td></tr>
+    </tbody>
+  </table>
+  <h2><span class="num">09</span>Riesgos clave</h2>
+  <table>
+    <thead><tr><th>Riesgo</th><th>Mitigación</th></tr></thead>
+    <tbody>
+      <tr><td>Flux.2 Klein no tiene API/pesos públicos cuando lo necesitas</td><td>Plan B: Flux.1-schnell o SDXL-Lightning. Pierdes posicionamiento sponsor pero idea sobrevive.</td></tr>
+      <tr><td>MiniCPM-V no identifica ingredientes mexicanos (chile poblano, nopales)</td><td>Few-shot en prompt; eventualmente fine-tune ligero del visión sobre 50 fotos etiquetadas</td></tr>
+      <tr><td>Flux.2 genera comida poco apetitosa</td><td>Itera prompts ("recipe magazine, warm light, top-down"); usa imagen final como ref para los pasos</td></tr>
+      <tr><td>Progress Validator da false positives</td><td>Conservador: solo dice "vas bien" si similitud es alta; default es "sigue" sin juicio fuerte</td></tr>
+      <tr><td>Latencia receta &gt; 30s</td><td>Streaming progresivo; paraleliza Flux + TTS</td></tr>
+      <tr><td>Modal cold start ~30-60s en Flux</td><td>Pre-warm 30s antes de filmar · <code>keep_warm=1</code> el día del demo</td></tr>
+      <tr><td>Persona del demo se quema/cocina mal</td><td>Practica la receta una vez antes · 2-3 candidatos de receta listos</td></tr>
+      <tr><td>Otro equipo presenta "recipe app con AI"</td><td>Diferéncialo con: closed-loop visual + español + cocina mx + dataset publicado + persona real</td></tr>
+    </tbody>
+  </table>
+  <h2><span class="num">10</span>Cómo gastar los créditos</h2>
+  <div class="grid-2">
+    <div class="card">
+      <h3>Modal · $250</h3>
+      <table>
+        <tr><td>Flux dev (días 1-9)</td><td>$5-15</td></tr>
+        <tr><td>Dataset cocina mx</td><td>$3-8</td></tr>
+        <tr><td>LoRA + sweeps</td><td>$4-5</td></tr>
+        <tr><td>Eval</td><td>$1</td></tr>
+        <tr><td>Inferencia grading jueces</td><td>$10-25</td></tr>
+        <tr><th>Subtotal</th><th>$25-65</th></tr>
+        <tr><th>+ Buffer</th><th>$30</th></tr>
+        <tr><th>Proyectado</th><th><strong>~$55-95 / $250</strong></th></tr>
+      </table>
+    </div>
+    <div class="card">
+      <h3>OpenAI Codex · $100</h3>
+      <table>
+        <tr><td>Codex CLI pair-programmer</td><td>$20-40</td></tr>
+        <tr><td>200 recetas mx sintéticas</td><td>$10-25</td></tr>
+        <tr><td>Prompts Flux por paso</td><td>$5-10</td></tr>
+        <tr><td>Reserva</td><td>$30</td></tr>
+        <tr><th>Proyectado</th><th><strong>~$65-105 / $100</strong></th></tr>
+      </table>
+    </div>
+  </div>
+  <div class="footnote">
+    <strong>Mantra del proyecto:</strong> "Una mamá cocinando frente a la cámara. Un platillo que se ve apetitoso. Una voz que la acompaña sin juzgar. Un paso a la vez."
+  </div>
+</div>
+</body>
+</html>

Strategy/estrategia.md ADDED Viewed

	@@ -0,0 +1,496 @@

+# Estrategia detallada — "Cocina Conmigo"
+> Documento de ejecución. Lee primero `plan.md` para el "qué" y el "por qué".
+> Este archivo es el "cómo": modelo mental, multi-agente, timeline, gasto de créditos, riesgos, snippets.
+---
+## 1. Modelo mental: la "receta" como objeto de estado
+La app no es un chatbot. Es una **máquina de estado** alrededor de un objeto `Recipe` que evoluciona en el tiempo. Ese estado se actualiza en cada turno.
+```python
+@dataclass
+class Recipe:
+    name: str                       # "Tinga de Pollo"
+    final_dish_image: bytes         # imagen Flux del plato final
+    available_ingredients: list[str]  # lo que la cámara vio en el refri
+    missing_ingredients: list[str]    # lo que falta + sus sustitutos
+    steps: list[Step]               # 5-7 pasos
+    current_step: int               # qué paso vamos haciendo
+    user_progress_photos: list[bytes]  # fotos que el usuario tomó
+@dataclass
+class Step:
+    n: int
+    instruction_text: str           # "Pica la cebolla en cubos chicos"
+    visual_target: bytes            # imagen Flux: "así debe verse el sartén"
+    duration_estimate: str          # "4 minutos"
+    audio_narration: bytes          # narración pre-renderizada
+    tip: str | None                 # "no la quemes"
+    tip_audio: bytes | None         # voz Cohere
+```
+Ventajas de pensarlo así:
+- Cada nodo del Workflow toma `Recipe` y devuelve `Recipe` modificada. Composable y observable.
+- El "replan" (no tengo cilantro) es una sola función `recipe.replan(missing="cilantro") → Recipe`.
+- El "validador" toma `Recipe` + `progress_photo` y devuelve `feedback`.
+---
+## 2. Los 5 agentes (multi-agente real, no simulado)
+| Agente | Responsabilidad | Trigger | Output |
+|---|---|---|---|
+| **Mise en Place** | Identificar ingredientes en foto del refri | foto del refri | `available_ingredients` |
+| **Recipe Planner** | Proponer 3 recetas factibles · armar la elegida | usuario elige idea | `Recipe` con steps |
+| **Step Illustrator** | Generar imagen-objetivo de cada paso + plato final | nueva receta | `Step.visual_target` para cada paso |
+| **Sous-Chef Narrator** | Narrar instrucciones por voz | paso activo | `Step.audio_narration` |
+| **Progress Validator** | Comparar foto del usuario vs imagen-objetivo | usuario sube foto mid-cooking | `feedback` (texto + voz tip) |
+Esto es un **sistema multi-agente real**: cada agente tiene su propia función, su propio modelo, y se comunican por estado compartido (`Recipe`). No es un agente único con tools — es 5 agentes en pipeline + closed-loop.
+> **Best Agent badge candidate.** Documenta esto en el README con un diagrama explícito.
+---
+## 3. El truco innovador: closed-loop visual
+```
+                ┌─────────────────────────────────────┐
+                │                                     │
+                ▼                                     │
+[Step Illustrator]──▶ visual_target ──▶ [Usuario cocina]
+                                              │
+                                              ▼
+                                    📸 progress_photo
+                                              │
+                                              ▼
+                                  [Progress Validator]
+                                    (MiniCPM-V)
+                                              │
+                          ┌───────────────────┤
+                          │                   │
+                    ✅ va bien           ❌ ajustar
+                          │                   │
+                    siguiente paso        [Recipe Planner]
+                                          replan/tip
+                                              │
+                                              └──────▶ vuelta al loop
+```
+Esta es **la innovación técnica** del proyecto. La mayoría de "recipe apps" son listas estáticas. Cocina Conmigo:
+1. Genera *visualmente* cómo debe verse cada paso (no solo texto).
+2. Acepta foto del usuario y la *compara* con el objetivo.
+3. Adapta el plan en vivo si algo no va.
+Sección dedicada en el README: *"How visual closed-loop cooking guidance works"*. Es también el blog post de Field Notes.
+---
+## 4. Cronograma — 10 días
+> ~50-70 horas de trabajo + 1 humano + Codex CLI como pair.
+### Día 1 — Setup + Modal Flux endpoint
+- `pip install gradio modal openai huggingface-hub diffusers llama-cpp-python`
+- `modal setup` y deploya el endpoint Flux que devuelve imagen dada un prompt.
+- Crea Space vacío en HF, push inicial.
+- **Entregable:** Space que muestra una imagen Flux dado un texto.
+### Día 2 — Vision: identificación de ingredientes
+- Carga MiniCPM-V Q4 GGUF en local.
+- Función: `identify_ingredients(fridge_photo) → list[str]`.
+- Prueba con 5 fotos de refri reales (el tuyo, el de tu mamá).
+- **Entregable:** dada foto del refri, devuelve lista correcta de 80%+ ingredientes visibles.
+### Día 3 — Recipe Planner LLM
+- Carga MiniCPM-4 Q4 GGUF.
+- Prompt template estructurado que devuelve JSON:
+  ```json
+  {
+    "name": "Tinga de Pollo",
+    "options": [{"name": "...", "why": "..."}, ...],
+    "steps": [{"n": 1, "instruction": "...", "duration": "...", "visual": "..."}],
+    "missing": ["cilantro"],
+    "substitutes": {"cilantro": ["perejil", "nada"]}
+  }
+  ```
+- Conecta Vision + Planner: foto refri → 3 opciones de receta.
+- **Entregable:** dada foto + selección, devuelve receta completa estructurada.
+### Día 4 — Step Illustrator (Flux.2 con consistencia)
+- Para cada `Step.visual` del JSON, llama Flux.2 endpoint con prompt:
+  > *"Top-down view of a kitchen pan with [step.visual]. Mexican cooking style. Warm lighting. Natural ingredients. Photorealistic, recipe magazine style."*
+- Para mantener estilo entre pasos: usa la imagen del paso anterior como `ref` con `strength=0.6` (más relajado que cuentos, porque el contenido cambia mucho).
+- Genera también imagen del plato final (sin `ref`).
+- **Entregable:** receta de 5 pasos cada uno con imagen-objetivo + foto del plato final.
+### Día 5 — Voz: narrador + tip-giver
+- **OpenBMB voice** para `Step.audio_narration`: instrucciones tono cálido y claro.
+- **Cohere Labs voice** para `Step.tip_audio`: tono más enérgico ("¡cuidado!").
+- Genera audio de los 5 pasos por adelantado (no en streaming, evita cold starts molestos).
+- **Entregable:** receta completa con narración audible.
+### Día 6 — UI Off-Brand: tarjeta de receta
+- `gr.Blocks` + CSS custom.
+- Layout: hero con imagen del plato final + título grande, abajo carrusel de pasos cada uno con `imagen objetivo + texto + botón "ya"`, modo cocina hands-free con texto enorme.
+- Estilo: serif elegante (`Lora`), paleta cálida tierra/dorado.
+- **Entregable:** Space que parece tarjeta de revista de cocina, no Gradio.
+### Día 7 — Gradio Workflows showcase
+- Reescribe pipeline como **Gradio Workflow** con nodos visibles.
+- Nodos: `📸 Fridge → 👁️ Vision → 🧠 Planner → 🎨 Illustrator → 🔊 Narrator → 📖 Recipe Card`.
+- Para `Progress Validator`, agrega rama: `📸 Progress Photo → 👁️ Validator → 💬 Feedback`.
+- Pestaña separada en el Space que muestra el grafo del Workflow corriendo en vivo.
+- **Entregable:** Workflow visualmente impresionante en pantalla. Diferenciador para jueces de Gradio.
+### Día 8 — Fine-tune del Planner en cocina mexicana
+- **Dataset sintético en Modal:** Codex API genera 200 recetas mexicanas en formato JSON estructurado (tinga, mole, chiles rellenos, sopes, pozole, etc.). Filtras manualmente las 150 mejores.
+- **LoRA en Modal A10G:** ~30-60 min de fine-tune sobre MiniCPM-4 4B.
+- **GGUF + push HF:** convierte a Q4_K_M, sube a HF Hub.
+- Reemplaza el Planner por la versión fine-tuneada.
+- **Entregable:** modelo `tu-usuario/cocinaconmigo-4b-mx-Q4_K_M-gguf` publicado.
+### Día 9 — STT + Progress Validator + eval
+- `faster-whisper tiny` en español: usuario pregunta hands-free.
+- Implementa **Progress Validator**: foto del usuario → MiniCPM-V compara contra `Step.visual_target` → genera feedback.
+- Eval: 10 recetas generadas, mide:
+  - % ingredientes correctamente identificados.
+  - % pasos con imagen-objetivo coherente.
+  - Calidad subjetiva de validación (5 fotos de progreso).
+- Sube traces al Hub (badge **Sharing is Caring**).
+- **Entregable:** app completa con voz IN, validador, traces publicados.
+### Día 10 — Demo video + README + blog + submit
+- **Filma a una persona real cocinando** una receta sugerida por la app, de principio a fin.
+- 60-90 segundos: foto del refri → 3 opciones → elige → cocina con voz → toma foto mid-cooking → app valida → plato final → la persona come.
+- README: badges declarados, diagrama, link al video, sección "How closed-loop visual cooking guidance works".
+- Blog post (badge **Field Notes**): "Le construí un sous-chef a mi mamá".
+- Submit + post social.
+---
+## 5. Decisiones técnicas explícitas
+### 5.1 Por qué Modal en runtime (rompiendo Off the Grid)
+Igual que en planes anteriores: Flux.2 9B en CPU del Space free es inviable (GB de RAM y minutos por imagen). Modal-powered es la elección obligada cuando el centro de la app es generación visual.
+### 5.2 Por qué cocina mexicana específicamente
+- Dataset acotado pero rico. Cubrible en 200 recetas.
+- Diferenciador cultural automático.
+- Se alinea con el público "para mi mamá" (si tu mamá es latina).
+- Si los jueces son mexicanos en Discord/Slack, +1.
+### 5.3 Por qué visual_target con Flux.2 en lugar de imagen stock
+- Stock photos tienen sesgo americano/europeo. Flux genera estilo mexicano si lo prompteas.
+- Stock no se ajusta al ingrediente exacto que tienes (Flux sí).
+- Esto es lo que hace única la app — es el wow factor.
+### 5.4 Por qué pre-renderizar audio en lugar de streaming
+- Latencia: streaming TTS tarda y se ve mal en demo.
+- Cocina es secuencial: sabes los 5 pasos cuando empieza la receta. Pre-render todo en paralelo.
+- Si el usuario hace replan, regeneras solo los pasos afectados.
+### 5.5 LoRA y no full fine-tune
+Mismo argumento de planes anteriores: 150-200 ejemplos = LoRA r=16 es suficiente. ~30 min A10G ≈ $1.
+### 5.6 Cómo gastar los $250 de Modal
+| Concepto | Estimado |
+|---|---|
+| Inferencia Flux.2 dev (días 1-9, ~5h GPU L4) | $5-15 |
+| Generación dataset sintético cocina mexicana (~2h) | $3-8 |
+| LoRA fine-tune + sweeps (~3h A10G) | $4-5 |
+| Eval pipeline | $1 |
+| Inferencia durante grading de jueces (~10h) | $10-25 |
+| **Subtotal** | **$25-65** |
+| Buffer | $30 |
+| **Total proyectado** | **~$55-95 / $250** |
+### 5.7 Cómo gastar los $100 de OpenAI Codex
+- Codex CLI durante 10 días como pair-programmer: $20-40.
+- Generación de 200 recetas mexicanas estructuradas (Día 8): $10-25.
+- Generación de prompts de Flux para los pasos (Día 4): $5-10.
+- Reserva: $30.
+---
+## 6. Riesgos y mitigaciones
+| Riesgo | Impacto | Mitigación |
+|---|---|---|
+| Flux.2 Klein no tiene API/pesos públicos cuando lo necesitas | Bloquea idea | Plan B: Flux.1-schnell o SDXL-Lightning. Pierdes tag sponsor pero idea sobrevive. |
+| MiniCPM-V no identifica ingredientes mexicanos (chile poblano, chayote, nopales) | Recipe Planner falla | Agrega few-shot examples al prompt; eventualmente fine-tune del visión sobre 50 fotos etiquetadas |
+| Flux.2 genera comida poco apetitosa/uncanny | Mata el demo | Itera prompts (style="recipe magazine, warm light, top-down"); usa imagen de plato final como ref para los pasos |
+| Latencia: receta completa tarda más de 30s en generarse | Demo aburrido | Streaming progresivo (muestra opción + plato final primero, pasos después); paraleliza Flux + TTS |
+| Modal cold start ~30-60s en Flux | Primera demo lenta | Pre-warm 30s antes de filmar; `keep_warm=1` el día del demo |
+| Validador de progreso da false positives ("vas bien" cuando no) | Confunde al usuario | Conservador: solo dice "vas bien" si la similitud es muy alta; default es "sigue" sin juicio fuerte |
+| TTS español sin acento mexicano | Suena raro | Si OpenBMB no tiene es-MX, usa Cohere o Kokoro con voz neutra; pre-graba para video |
+| Usuario del demo cocina mal/se quema | Mata el video | Practica la receta una vez antes de filmar; ten 2-3 candidatos de receta listos |
+| Otro equipo presenta "recipe app con AI" | Compite por premios | Diferénciate con: closed-loop visual + español + cocina mexicana específica + dataset publicado + persona real cocinando + Workflow visible |
+| Workflows de Gradio inestable (lanzado ayer) | Rompe app | Versión sin Workflows como backup. Workflows es decoración. |
+---
+## 7. Plan B — corte de scope
+Si en Día 7 ves que no llegas, recorta features en este orden:
+| # | Cortar | Pierdes | Conservas |
+|---|---|---|---|
+| 1 | STT (preguntas hands-free por voz) | comodidad demo | input por texto + foto |
+| 2 | 2da voz (Cohere tip-giver) | un sponsor de voz | narrador único |
+| 3 | Progress Validator (closed-loop) | **Best Agent badge** + innovación principal | demo lineal sin loop |
+| 4 | Fine-tune del Planner | **Well-Tuned badge** | base model + prompting |
+| 5 | Gradio Workflows showcase | diferenciador "fresh" | pipeline Python |
+| 6 | UI super-custom | **Off-Brand badge** | UI default |
+**NUNCA cortar:**
+- Vision + Planner + Step Illustrator + Narrator + UI mínima + video con persona real cocinando.
+Eso solo ya entra fuerte a Backyard AI track.
+---
+## 8. Métricas de éxito (auto-evaluación pre-submit)
+Antes de mandar:
+- [ ] Una persona real cocinó una receta entera con la app y se la comió.
+- [ ] El video tiene una cara humana y un plato terminado en al menos 30s de los 90s.
+- [ ] La app identifica correctamente ≥4 de 5 ingredientes en una foto típica de refri.
+- [ ] Las imágenes de Flux para los pasos se ven *apetitosas* (test: si las muestras a alguien sin contexto, dice "se ve rico").
+- [ ] Una receta completa se genera en menos de 30s (texto + 5 imágenes + audio).
+- [ ] El Progress Validator funciona en al menos 5 de 10 fotos de progreso reales.
+- [ ] El README tiene un diagrama y la sección "How closed-loop cooking works".
+- [ ] Hay 3 recetas pre-renderizadas listas para que jueces las vean sin esperar.
+- [ ] Total params declarado y verificado ≤ 32B.
+- [ ] Sin secrets hardcoded.
+Si fallas más de 2, no submitas; arregla.
+---
+## 9. Lo que NO debes hacer
+- **No** intentes generar video del platillo. Imagen estática se ve mejor que video AI mediocre.
+- **No** hagas más de 7 pasos por receta. Atención del juez = 60-90s.
+- **No** soportes 100 recetas. Soporta 20 recetas mexicanas excelentes y di "más recetas pronto".
+- **No** subas fotos del refri real con productos identificables (marcas, info personal). Borra labels.
+- **No** persigas Off the Grid. Decisión ya tomada.
+- **No** dejes el video de demo para el último día sin practicar la receta antes.
+- **No** publiques tokens en el repo.
+- **No** generes recetas con ingredientes raros que la mayoría no tenga (cocina accesible > cocina chef).
+---
+## 10. Pitch del README (esqueleto)
+```markdown
+# Cocina Conmigo
+> A visual sous-chef that sees what's in your fridge,
+> shows you what each step should look like, and walks you through it
+> with voice — hands-free.
+[60-second demo video embed: tu mamá cocinando tinga]
+## Why it shouldn't exist (but does)
+Every recipe app is a list of steps. Cocina Conmigo is a closed-loop assistant:
+it generates the *target image* of each cooking step with Flux.2, listens
+when you ask "¿voy bien?", and adapts when you say "no tengo cilantro."
+## Tech
+- 👁️ MiniCPM-V — sees your fridge + validates your progress
+- 🧠 MiniCPM-4 4B (LoRA fine-tuned on Mexican cuisine) — recipe planner
+- 🎨 Flux.2 Klein 9B (Modal endpoint) — generates target images per step
+- 🔊 OpenBMB voice — sous-chef narrator
+- 🎭 Cohere voice — tip-giver second voice
+- 🎙️ Whisper-tiny — voice input
+- ⚙️ Gradio Workflows — visible pipeline of nodes
+Total params: ~17B (≤ 32B ✓)
+## Badges
+✓ Llama Champion · ✓ Well-Tuned · ✓ Off-Brand · ✓ Sharing is Caring · ✓ Field Notes
+## Built for
+My mom. She makes great mole. She can never remember tinga.
+## Try it
+[HF Space link]
+```
+---
+## 11. Apéndice: snippets clave
+### 11.1 Mise en Place agent (vision)
+```python
+def identify_ingredients(image: PIL.Image) -> list[str]:
+    prompt = """Veo esta foto de un refrigerador o despensa.
+    Lista TODOS los ingredientes que se ven, en español, en JSON:
+    {"ingredients": ["pollo", "cebolla", "cilantro", ...]}
+    Solo ingredientes alimentarios, no contenedores."""
+    out = mini_cpm_v.create_chat_completion(messages=[
+        {"role": "user", "content": [
+            {"type": "image_url", "image_url": pil_to_data_url(image)},
+            {"type": "text", "text": prompt}
+        ]}
+    ])
+    return json.loads(out["choices"][0]["message"]["content"])["ingredients"]
+```
+### 11.2 Recipe Planner agent (LLM)
+```python
+SYS = """Eres un chef mexicano. Generas recetas a partir de ingredientes
+disponibles. Prefiere cocina mexicana tradicional, accesible.
+Salida JSON estricta:
+{
+  "name": "...",
+  "options": [{"name": "...", "why": "..."}],
+  "steps": [
+    {"n": 1, "instruction": "...", "duration": "4 min",
+     "visual": "english visual description for image gen",
+     "tip": "optional warning or tip"}
+  ],
+  "missing": ["cilantro"],
+  "substitutes": {"cilantro": ["perejil", "nada"]},
+  "final_dish_visual": "english visual description of the final plated dish"
+}
+"""
+def plan_recipe(ingredients, choice=None):
+    msgs = [{"role": "system", "content": SYS}]
+    msgs.append({"role": "user", "content":
+        f"Tengo: {', '.join(ingredients)}.\n"
+        + (f"Quiero hacer: {choice}." if choice else "Propón 3 opciones.")})
+    raw = llm.create_chat_completion(messages=msgs, temperature=0.7)
+    return json.loads(raw["choices"][0]["message"]["content"])
+```
+### 11.3 Step Illustrator (Flux endpoint)
+```python
+import modal
+app = modal.App("cocina-flux")
+image = modal.Image.debian_slim().pip_install("torch","diffusers","transformers","accelerate","Pillow")
+@app.cls(image=image, gpu="L4", scaledown_window=180, keep_warm=0)
+class FluxKlein:
+    @modal.enter()
+    def load(self):
+        from diffusers import FluxPipeline
+        self.pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.2-klein",
+                                                 torch_dtype="bfloat16").to("cuda")
+    @modal.method()
+    def render_step(self, visual: str, ref_png: bytes | None = None) -> bytes:
+        from PIL import Image; import io
+        prompt = (f"Top-down photo of a kitchen pan or plate showing {visual}. "
+                  f"Mexican home cooking, warm natural lighting, recipe magazine "
+                  f"style, photorealistic, appetizing.")
+        if ref_png:
+            ref = Image.open(io.BytesIO(ref_png)).convert("RGB")
+            img = self.pipe(prompt=prompt, image=ref, strength=0.6,
+                            num_inference_steps=4).images[0]
+        else:
+            img = self.pipe(prompt=prompt, num_inference_steps=4).images[0]
+        buf = io.BytesIO(); img.save(buf, "PNG"); return buf.getvalue()
+```
+### 11.4 Progress Validator (closed-loop)
+```python
+def validate_progress(target_image: PIL.Image, user_photo: PIL.Image,
+                      step_instruction: str) -> dict:
+    prompt = f"""Compara estas dos fotos de cocina.
+    Foto 1 (objetivo): cómo debe verse después del paso "{step_instruction}".
+    Foto 2 (usuario): cómo va el usuario.
+    Responde en JSON:
+    {{"verdict": "go|wait|fix", "feedback_es": "...", "tip": "..." | null}}
+    - "go": va bien, siguiente paso
+    - "wait": le falta tiempo
+    - "fix": algo se ve mal, sugiere ajuste
+    """
+    out = mini_cpm_v.create_chat_completion(messages=[
+        {"role": "user", "content": [
+            {"type": "image_url", "image_url": pil_to_data_url(target_image)},
+            {"type": "image_url", "image_url": pil_to_data_url(user_photo)},
+            {"type": "text", "text": prompt}
+        ]}
+    ])
+    return json.loads(out["choices"][0]["message"]["content"])
+```
+### 11.5 UI Off-Brand (recipe card)
+```python
+import gradio as gr
+CSS = """
+@import url('https://fonts.googleapis.com/css2?family=Lora:wght@400;700&family=Inter:wght@400;600&display=swap');
+.gradio-container {background: #f5ecd9 !important; font-family: 'Inter', sans-serif !important;}
+.recipe-hero {background: #fffbf0; border-radius: 14px; padding: 28px;
+  box-shadow: 0 8px 24px rgba(0,0,0,0.12); border: 1px solid #d8c9ad;}
+.recipe-hero h1 {font-family: 'Lora', serif !important; font-size: 36px !important;
+  margin: 0 0 6px !important; color: #6b4a2a !important;}
+.step-card {background: #fffbf0; border-left: 4px solid #a85c2a;
+  border-radius: 8px; padding: 18px 22px; margin: 12px 0;}
+.step-card h3 {font-family: 'Lora', serif !important; margin: 0 !important;}
+.step-card p {font-size: 17px !important; line-height: 1.6;}
+button.primary {background: #a85c2a !important; font-family: 'Inter', sans-serif !important;
+  font-weight: 600 !important; font-size: 16px !important; padding: 14px 22px !important;}
+"""
+with gr.Blocks(css=CSS, title="Cocina Conmigo") as demo:
+    gr.Markdown("# 👩‍🍳 Cocina Conmigo")
+    fridge = gr.Image(label="📸 Foto de tu refri o despensa", type="pil")
+    btn = gr.Button("¿Qué cocino?", variant="primary")
+    with gr.Column(elem_classes=["recipe-hero"]):
+        title = gr.Markdown()
+        final_img = gr.Image(show_label=False)
+        steps_box = gr.Column()
+    progress = gr.Image(label="📸 Tómame foto de tu progreso", type="pil")
+    feedback = gr.Markdown()
+    # callbacks omitidos
+```
+### 11.6 LoRA fine-tune del Planner en Modal
+```python
+@app.function(image=image_train, gpu="A10G", timeout=60*60*2,
+              volumes={"/cache": modal.Volume.from_name("hf-cache", create_if_missing=True)})
+def train_planner():
+    import os; os.environ["HF_HOME"] = "/cache"
+    from transformers import AutoModelForCausalLM, AutoTokenizer
+    from peft import LoraConfig, get_peft_model
+    from trl import SFTTrainer, SFTConfig
+    from datasets import load_dataset
+    base = "openbmb/MiniCPM-4-Base"
+    tok = AutoTokenizer.from_pretrained(base, trust_remote_code=True)
+    model = AutoModelForCausalLM.from_pretrained(base, trust_remote_code=True,
+                                                 device_map="cuda", torch_dtype="bfloat16")
+    model = get_peft_model(model, LoraConfig(r=16, lora_alpha=32,
+                                              target_modules="all-linear"))
+    ds = load_dataset("tu-usuario/recetas-mexicanas-sft", split="train")
+    SFTTrainer(model=model, tokenizer=tok, train_dataset=ds,
+               args=SFTConfig(output_dir="/cache/out", num_train_epochs=2,
+                              per_device_train_batch_size=4, learning_rate=2e-4,
+                              push_to_hub=True,
+                              hub_model_id="tu-usuario/cocinaconmigo-4b-mx")
+               ).train()
+```
+---
+## 12. Lectura recomendada antes del Día 1
+- `Context/guia-tecnologias.md` (sección 3 Modal, sección 4 llama.cpp).
+- HF Black Forest Labs: <https://huggingface.co/black-forest-labs> — confirma versión Flux.2 Klein.
+- HF MiniCPM-V: <https://huggingface.co/openbmb> — versión vision con GGUF.
+- Modal stable-diffusion example: <https://github.com/modal-labs/modal-examples/tree/main/06_gpu_and_ml/stable_diffusion>.
+- Diffusers img2img: <https://huggingface.co/docs/diffusers/using-diffusers/img2img>.
+- Gradio Workflows: <https://www.gradio.app/guides> (busca el guide más reciente).
+- Cohere Labs voice: confirma con sponsor el modelo exacto disponible.
+> Cocina con tu mamá una vez antes de empezar a programar. Te va a aclarar más sobre qué necesita tu app que cualquier brainstorm. Suerte.

Strategy/plan.md ADDED Viewed

	@@ -0,0 +1,245 @@

+# Plan ganador — "Cocina Conmigo"
+> Un sous-chef multimodal que ve lo que tienes en el refri, te dice qué cocinar, te muestra cómo debe verse cada paso con Flux.2, y te narra todo por voz mientras cocinas con las manos llenas.
+>
+> Hackathon "Small models / Big adventures" — junio 2026.
+---
+## TL;DR
+**Idea elegida:** **Cocina Conmigo** — un copiloto de cocina hands-free que combina visión, razonamiento, generación de imagen en tiempo real, y voz, para acompañarte de principio a fin: desde *"¿qué cocino con esto?"* hasta *"¿voy bien?"*.
+**Por qué esta y no otra:** es la única idea que **(1) está fuera de las 11 ideas pre-cocinadas por OpenBMB**, **(2) usa Flux.2 + voces + Workflows como núcleo**, y **(3) tiene utilidad real, diaria y universal**. Nadie cocina como hobby; todos cocinan por necesidad.
+---
+## Por qué cambió el plan respecto a iteraciones anteriores
+| Iteración | Idea | Por qué se descartó |
+|---|---|---|
+| v1 | Abuelita (parent phone helper) | **Está en la lista pre-cocinada de OpenBMB para Backyard AI.** 5-15 equipos van a hacer la misma cosa. |
+| v2 | Cuentacuentos (storyteller ilustrado) | **Está en la lista pre-cocinada de OpenBMB para Thousand Token Wood ("voice storyteller").** Mismo problema de saturación. |
+| v3 (ésta) | **Cocina Conmigo** | Refinamiento de **tu propia idea #1**, ahora viable de verdad gracias a Flux.2. **No está en ninguna lista pre-cocinada.** |
+La regla estratégica: **usar los modelos de los sponsors, no copiar sus templates de proyecto.**
+---
+## Las 12 ideas en zona prohibida (clúster OpenBMB)
+| Backyard AI | Thousand Token Wood |
+|---|---|
+| Parent phone helper | Voice storyteller |
+| Receipt / bill explainer | Visual mystery box |
+| Shop menu / repair manual | AI museum |
+| Offline personal assistant / voice companion | Doodle creature |
+| | Dream postcard gen |
+| | Omni-modal adventure |
+| | Tiny local NPC / character agent |
+Y de tus 5 ideas originales, también caen:
+- #3 cortes de cabello (tú mismo dijiste "ya está muy trabajado")
+- #4 museum Q&A (choca con "AI museum")
+**Quedan vivas, fuera de zona prohibida:**
+- #1 Recetas (→ **Cocina Conmigo**, esta propuesta)
+- #2 Detector de intenciones (no usa Flux.2, demo aburrida)
+- #5 Outfits con armario (alternativa B, ver final del documento)
+---
+## El producto en una frase
+> *"Mi mamá me pidió que le enseñara a hacer ramen. Le construí un sous-chef que vive en su tablet."*
+---
+## Las 4 historias del demo
+### 1. *"Tengo esto en el refri"*
+```
+👩 Mamá toma foto del refri abierto.
+🤖 [MiniCPM-V] "Veo: pollo, cebolla, jitomate, cilantro, tortillas, queso."
+🤖 [LLM] "Te puedo proponer: tinga de pollo, enchiladas, o quesadillas. ¿Qué traes ganas?"
+👩 "Tinga."
+🤖 [Flux.2] genera foto del platillo final, hermosa, mexicana.
+🤖 "Perfecto. Te tomará 35 minutos. ¿Empezamos?"
+```
+### 2. *"Cocina paso a paso"* (hands-free)
+```
+🤖 [Flux.2] muestra: olla con cebolla acitronándose
+🤖 [Voz OpenBMB] "Pica la cebolla en cubitos chicos y ponla en aceite caliente."
+👩 (cocinando, manos sucias)
+👩 "¿Cuánto tiempo?"
+🤖 [Voz] "Hasta que se vea transparente. Como 4 minutos."
+```
+### 3. *"¿Voy bien?"* (visión en loop)
+```
+👩 (toma foto del sartén con cebolla)
+🤖 [MiniCPM-V] compara contra imagen objetivo.
+🤖 [Voz Cohere — el "tip-giver"] "Le falta un poquito. Súbele 1 minuto más, está bien."
+```
+### 4. *"No tengo cilantro"* (replan adaptativo)
+```
+👩 "No tengo cilantro."
+🤖 [LLM] re-planea sobre la marcha.
+🤖 [Voz] "No pasa nada. Le ponemos perejil o nada. Sigue siendo tinga."
+🤖 [Flux.2] regenera la foto del plato final, ahora sin cilantro.
+```
+Las 4 historias usan los **mismos 5 modelos**. Una sola pipeline.
+---
+## Por qué este plan **gana** este hackathon
+### 1. "Build for someone you actually know" → Backyard AI track
+La descripción literal del track dice: *"Solve a real problem for someone you actually know. Pick a person — a neighbor, a parent, a small-business owner..."*. Tu mamá. Tu hermana. Tu hermano que vive solo. **Todos** cocinan. Pocas apps de hackathon van a tener un usuario tan cercano y tan recurrente.
+### 2. Aprovecha **todos** los assets sponsor sin copiar templates
+| Asset | Cómo se usa |
+|---|---|
+| **Flux.2 Klein 9B** (sponsor) | Genera la imagen-objetivo del platillo + "esto debes ver" en cada paso · i2i para ajustes |
+| **MiniCPM-V** (OpenBMB) | Visión: identifica ingredientes + valida progreso ("¿voy bien?") |
+| **MiniCPM razonamiento** (OpenBMB) | Recipe Planner: arma receta + replan adaptativo |
+| **OpenBMB voice / TTS** | Voz principal del sous-chef (cálida, paciente) |
+| **Cohere Labs voice** (sponsor) | Segunda voz: tips, advertencias ("¡cuidado, se quema!") |
+| **Whisper-tiny** | STT: preguntas hands-free mientras cocinas |
+| **Gradio Workflows** | UI de nodos visible: Vision → Planner → Illustrator → Narrator → Validator |
+| **Modal $250** | Hostea Flux.2 en GPU + dataset sintético + LoRA fine-tune |
+| **OpenAI Codex $100** | Pair-programmer y generador de dataset de recetas |
+Todos los sponsors tocados. Cero ideas copiadas.
+### 3. **Innovación técnica concreta**: el bucle visual cerrado
+La mayoría de "recipe apps" del mundo son listas de pasos. Cocina Conmigo introduce un **closed-loop visual**:
+```
+[Flux.2 muestra paso ideal]  ──▶  [Usuario cocina]
+              ▲                          │
+              │                          ▼
+[LLM ajusta plan]  ◀──  [MiniCPM-V valida foto del usuario]
+```
+Esto es un agente real, no un wrapper. Best Agent badge en juego.
+### 4. Demo apetitoso = video viral
+Persona real cocinando + voz cálida + ilustraciones live + "¡me quedó igual!" + plato final que se come frente a la cámara. Best Demo + Community Choice por inercia. **Nadie va a recordar la submission #14 de "voice storyteller"; van a recordar el video donde tu mamá hace tinga con AI.**
+### 5. Diferenciación cultural sostenible
+- **Español-mexicano-first** — diferenciador en hackathon US-céntrico.
+- **Cocina mexicana** como dataset de fine-tune — territorio que pocos van a tocar.
+- "Para mi mamá" como historia: emocional + universal.
+---
+## Arquitectura (resumen — ver `arquitectura.html`)
+5 nodos en un Gradio Workflow visible:
+```
+   [📸/🎙️ Input]  ──▶  [👁️ Vision MiniCPM-V]  ──▶  [🧠 Recipe Planner]  ──▶  [🎨 Step Illustrator Flux.2]
+                                                          │
+                                                          ▼
+                                                   [🔊 Sous-Chef Narrator OpenBMB]  +  [🎭 Tip-Giver Cohere]
+                                                          │
+                                                          ▼
+                                                  [✅ Progress Validator]  ──▶  loop al usuario
+```
+| Nodo | Modelo | Tamaño | Rol |
+|---|---|---|---|
+| Vision In | MiniCPM-V 2.6 / 4 (Q4 GGUF) | ~2-4B | Identifica ingredientes + valida progreso |
+| Planner | MiniCPM-4 4B (LoRA en cocina mexicana) | ~4B | Genera receta JSON estructurado · replan |
+| Illustrator | Flux.2 Klein 9B (Modal GPU) | 9B | Imagen final + paso-a-paso, i2i para consistencia |
+| Narrator | OpenBMB voice / Kokoro | ~1B | Voz principal: instrucciones |
+| Tip-Giver | Cohere Labs voice | ~1B | Segunda voz: warnings, encouragement |
+| STT (opcional) | Whisper-tiny | ~40M | "¿voy bien?" "¿cuánto tiempo?" |
+**Total: ~17B parámetros** (cap 32B ✓)
+**Donde corre:**
+- Vision, Planner, voces, STT → CPU del HF Space (llama.cpp + bindings ligeros)
+- **Flux.2 → endpoint Modal con GPU L4** (no aguanta CPU del Space)
+> Mismo tradeoff que los planes anteriores: **rompemos Off the Grid** intencionalmente para preservar calidad de imagen y latencia. A cambio calificamos para Modal Awards.
+---
+## Badges objetivo (5/6)
+| Badge | Cómo |
+|---|---|
+| ✓ **Llama Champion** | Vision + Planner via `llama-cpp-python` con GGUF Q4 |
+| ✓ **Well-Tuned** | LoRA del Planner en dataset de cocina mexicana, publicado en HF |
+| ✓ **Off-Brand** | UI estilo "tarjeta de receta" + modo cocina hands-free, no parece Gradio default |
+| ✓ **Sharing is Caring** | Dataset de recetas mexicanas + agent traces + recetas generadas, todo al Hub |
+| ✓ **Field Notes** | Blog: "Le construí un sous-chef a mi mamá" |
+| ✗ **Off the Grid** | Sacrificio consciente: Flux.2 corre en Modal |
+5 badges + Modal-powered fuerte = competitivo para **Bonus Quest Champion ($2K)**.
+---
+## Premios objetivo (proyección)
+| Premio | Probabilidad | Por qué |
+|---|---|---|
+| **Backyard AI Track** ($1K–$4K) | **Alta** | Idea es texto literal del track. Demo emocional. |
+| **Modal Awards** ($3K–$10K credits) | **Alta** | Flux en Modal en runtime + entrenamiento offline. Modal-powered de manual. |
+| **OpenBMB Award** ($1K–$2.5K) | **Alta** | Usa modelos OpenBMB en 3 roles (vision, planner, voice) sin copiar template |
+| **Best Demo** ($1K) | **Alta** | Persona cocinando + comida final + voz = video apetitoso |
+| **Community Choice** ($2K) | **Alta** | Apela a memoria emocional universal (tu mamá cocinando) |
+| **Bonus Quest Champion** ($2K) | Media-alta | 5/6 badges es competitivo |
+| **Best Agent** ($1K) | Media-alta | Closed-loop multi-agent real (5 agentes) |
+| **Off-Brand** ($1.5K) | Media | UI tarjeta-de-receta tiene buenas chances |
+| **Tiny Titan** ($1.5K) | Baja | Flux.2 9B nos saca del rango ≤4B |
+**Cota razonable acumulada:** $5K–$12K cash + $3K–$10K Modal credits.
+---
+## Las 3 condiciones que pone Idea.md
+| Condición | Cómo se cumple |
+|---|---|
+| **Innovador** | Closed-loop visual (Flux genera ideal → usuario cocina → vision valida → planner ajusta) — no existe en apps de receta |
+| **Fresco** | Combina Flux.2 (nuevo) + Workflows (lanzado ayer) + voces multi-sponsor + cocina hands-free. Ninguna submission tendrá esa combinación. |
+| **Útil** | Cocinar es diario, universal, recurrente. La app reemplaza Google + YouTube + adivinar. |
+---
+## Decisiones que tienes que tomar tú
+| Decisión | Recomendación |
+|---|---|
+| ¿Cocina Conmigo o Mi Espejo (outfits)? | **Cocina.** Menor riesgo técnico (Flux generando platos > generando personas reales con ropa). Más universal. |
+| ¿Cocina mexicana o cocina general? | **Mexicana.** Diferenciador + fine-tune en dataset acotado y rico. |
+| ¿Persona real para el demo? | **Sí, no negociable.** Tu mamá, tu pareja, tu vecina. Que coma frente a la cámara al final. |
+| ¿Empiezas con texto o con voz/foto? | **Empieza con foto del refri + texto.** Voz se agrega en Día 7-9. |
+| ¿Cuántos pasos por receta? | 5-7 pasos. Más es muy largo para el demo, menos no es una receta. |
+---
+## Plan B — alternativa "Mi Espejo"
+Si por cualquier razón Cocina Conmigo no avanza (ej. Flux.2 genera platillos feos consistentemente), pivota a **"Mi Espejo"** (refinamiento de tu idea #5):
+- 📸 Subes foto tuya + fotos del armario.
+- 🧠 Stylist LLM combina outfits según ocasión + tendencia.
+- 🎨 **Flux.2 i2i te genera vistiendo cada combinación.**
+- 🔊 Voz comenta el look.
+Mismas badges, mismo track (Backyard), pero más alto wow visual y más alto riesgo (uncanny valley con personas reales). **Es plan B**, no plan A.
+---
+## Siguiente paso
+Lee **`estrategia.md`** (timeline 10 días, gasto Modal/Codex, riesgos+mitigaciones, snippets) y **`arquitectura.html`** (diagrama del sistema + las 4 historias del demo + Workflow visual). Luego abre Codex CLI y haz el "hola mundo" del Día 1: un endpoint Modal que devuelve una imagen Flux.2 de un platillo dado un nombre de receta.
+> *"Cocinar es la última cosa que la IA debería poder ayudarte a hacer bien. Y por eso es la mejor cosa que puedes ganar haciendo."*

Strategy/plan_implementacion.md ADDED Viewed

	@@ -0,0 +1,674 @@

+# Implementation Plan — "Cook With Me"
+> Step-by-step implementation guide for developers building the multimodal cooking sous-chef Gradio app for Hugging Face Spaces.
+>
+> **Hackathon:** Small models / Big adventures — June 2026
+> **Read first:** `plan.md` (the *what* and *why*) and `estrategia.md` (the *how* at a strategic level). This document is the *how* at a tactical level — turn this into code.
+---
+## 0. Locked decisions (do not re-discuss)
+| Decision | Value | Reason |
+|---|---|---|
+| UI framework | **Gradio** | Hackathon requirement |
+| Hosting | **Hugging Face Space** | Hackathon requirement |
+| Inference runtime (text + vision) | **llama.cpp** via `llama-cpp-python` | Runs inside the Space CPU, no external APIs needed for now. Future: migrate to Modal |
+| Image generation | **FLUX.2 Klein 9B** (`black-forest-labs/FLUX.2-klein-9B`) | Sponsor model; runs in the Space if a GPU Space is rented (or via `enable_model_cpu_offload()` as fallback). Plan to migrate this specific component to Modal post-hackathon |
+| Recipe planner / reasoning | **`openbmb/MiniCPM-V-4`** (GGUF) | Provided requirement |
+| Vision (ingredient ID + progress validator) | **`openbmb/MiniCPM-V-4.6`** (GGUF) | Provided requirement |
+| Text-to-speech | **OpenBMB VoxCPM2** | Provided requirement |
+| Recipe dataset | **`thedevastator/better-recipes-for-a-better-life`** (Kaggle) — international cuisine | Provided requirement; not limited to Mexican food |
+| App language | **English only** | Provided requirement |
+| Final output | **Recipe + step images + voice + nutritional values** | Provided requirement |
+| External API calls at runtime | **None** | "llama.cpp inside the Space" mandate |
+---
+## 1. Architecture (final, English-only, llama.cpp-first)
+```
+                          ┌──────────────────────────────────────┐
+                          │     Hugging Face Space (Gradio)      │
+                          │   (CPU + optional GPU upgrade)       │
+                          ├──────────────────────────────────────┤
+   📸 Fridge photo  ─────▶│  [Vision Agent]                      │
+                          │   MiniCPM-V-4.6 GGUF (llama.cpp)     │
+                          │   → list[ingredient]                  │
+                          │              │                        │
+                          │              ▼                        │
+   🥘 User picks dish ───▶│  [Recipe Planner]                    │
+                          │   MiniCPM-V-4 GGUF (llama.cpp)       │
+                          │   + retrieval over Kaggle dataset    │
+                          │   → Recipe JSON (steps, nutrition)   │
+                          │              │                        │
+                          │              ▼                        │
+                          │  [Step Illustrator]                   │
+                          │   FLUX.2 Klein 9B (diffusers)        │
+                          │   → PNG per step + final dish        │
+                          │              │                        │
+                          │              ▼                        │
+                          │  [Narrator]                           │
+                          │   VoxCPM2 → MP3 per step             │
+                          │              │                        │
+                          │              ▼                        │
+   📸 Progress photo ────▶│  [Progress Validator]                │
+                          │   MiniCPM-V-4.6 (vision compare)     │
+                          │   → "go / wait / fix" + tip          │
+                          └──────────────────────────────────────┘
+```
+**Total parameter count (≤ 32B requirement):**
+- MiniCPM-V-4 (reasoning) ≈ 4B
+- MiniCPM-V-4.6 (vision) ≈ 4.6B
+- FLUX.2 Klein ≈ 9B
+- VoxCPM2 ≈ 1B (estimate)
+- **Total ≈ 18.6B ✓**
+---
+## 2. Repository layout
+```
+cook-with-me/
+├── app.py                      # Gradio entrypoint (Space looks for this)
+├── requirements.txt
+├── packages.txt                # apt packages (ffmpeg, libsndfile1)
+├── README.md                   # Space card (HF requires YAML frontmatter)
+├── .gitignore
+├── src/
+│   ├── __init__.py
+│   ├── config.py               # paths, model IDs, constants
+│   ├── models/
+│   │   ├── __init__.py
+│   │   ├── vision.py           # MiniCPM-V-4.6 wrapper (llama-cpp)
+│   │   ├── planner.py          # MiniCPM-V-4 wrapper (llama-cpp)
+│   │   ├── illustrator.py      # FLUX.2 Klein wrapper (diffusers)
+│   │   ├── narrator.py         # VoxCPM2 wrapper
+│   │   └── loader.py           # lazy singletons + GGUF download
+│   ├── agents/
+│   │   ├── mise_en_place.py    # ingredient identification
+│   │   ├── recipe_planner.py   # builds Recipe object
+│   │   ├── step_illustrator.py # per-step image gen
+│   │   ├── narrator.py         # per-step TTS
+│   │   └── progress_validator.py
+│   ├── data/
+│   │   ├── recipe_index.py     # loads Kaggle dataset, builds retrieval
+│   │   └── nutrition.py        # USDA-style nutrition computation
+│   ├── pipeline.py             # Recipe state machine, orchestration
+│   ├── prompts/
+│   │   ├── vision_prompt.txt
+│   │   ├── planner_system.txt
+│   │   └── validator_prompt.txt
+│   └── ui/
+│       ├── theme.py            # custom CSS (Off-Brand badge)
+│       └── components.py       # reusable Gradio Blocks pieces
+├── scripts/
+│   ├── download_models.py      # pre-warms GGUF + Flux weights at build time
+│   ├── build_recipe_index.py   # caches Kaggle dataset locally
+│   └── smoke_test.py           # end-to-end validation before push
+└── assets/
+    ├── sample_fridge_1.jpg
+    └── sample_progress_1.jpg
+```
+---
+## 3. Phase-by-phase plan (10 days)
+> Each phase has: **goal**, **tasks**, **deliverable**, **verification check**. Do not move to the next phase if verification fails.
+---
+### Phase 0 — Day 0 (½ day): Account + tooling setup
+**Goal:** every credential and CLI is ready before writing code.
+**Tasks**
+1. Create or confirm Hugging Face account; generate a **write token** (Settings → Access Tokens). Store as `HF_TOKEN` env var locally.
+2. Install Hugging Face CLI: `pip install -U huggingface_hub` then `huggingface-cli login`.
+3. Install Kaggle CLI: `pip install kaggle`. Place `kaggle.json` (Account → API → Create New Token) in `~/.kaggle/kaggle.json` with `chmod 600`.
+4. Install OpenAI Codex CLI (pair-programmer) and verify your $100 credit is active.
+5. Install local Python 3.11 venv: `python -m venv .venv && source .venv/bin/activate`.
+6. Create the repo locally: `git init cook-with-me && cd cook-with-me`.
+7. Create an empty Hugging Face Space: huggingface.co → New Space → SDK = **Gradio**, Hardware = **CPU basic** (upgrade later if you need GPU for FLUX). Clone it and copy your repo skeleton into it.
+8. Verify model availability: open in a browser and confirm pages exist:
+   - `huggingface.co/openbmb/MiniCPM-V-4`
+   - `huggingface.co/openbmb/MiniCPM-V-4-6`
+   - `huggingface.co/openbmb/VoxCPM2` (or whatever the exact repo name is — search "VoxCPM" on HF)
+   - `huggingface.co/black-forest-labs/FLUX.2-klein-9B`
+**Deliverable:** empty Space deployed showing "Hello World" Gradio.
+**Verify:** `https://huggingface.co/spaces/<you>/cook-with-me` loads.
+---
+### Phase 1 — Day 1: Project skeleton + recipe dataset ingestion
+**Goal:** the Kaggle dataset is downloaded, parsed, and cached as a local artifact ready for retrieval.
+**Tasks**
+1. Write `requirements.txt` (initial version — packages will be added as phases progress):
+   ```text
+   gradio>=4.44
+   huggingface_hub>=0.24
+   llama-cpp-python>=0.3.2
+   numpy
+   pandas
+   Pillow
+   pydantic>=2
+   sentence-transformers
+   ```
+2. Write `packages.txt`:
+   ```text
+   ffmpeg
+   libsndfile1
+   ```
+3. Write `scripts/build_recipe_index.py`:
+   - Use `kagglehub.load_dataset(KaggleDatasetAdapter.PANDAS, "thedevastator/better-recipes-for-a-better-life", file_path)` — discover `file_path` by listing the dataset files first via `kagglehub.dataset_download`.
+   - Normalize columns: `name`, `ingredients` (list[str]), `instructions` (list[str]), `cuisine` (str if present, else "international"), `prep_time`, `servings`.
+   - Drop rows missing critical fields. Lowercase + strip ingredient strings.
+   - Save to `data/recipes.parquet` (~5–50MB depending on dataset size).
+   - Build sentence embeddings of the recipe **name + first 3 ingredients** using `sentence-transformers/all-MiniLM-L6-v2` and save to `data/recipes_emb.npy`.
+   - This script runs **once locally**; commit the parquet + npy files to the repo (or to a private HF Dataset, then download in `app.py`). If files exceed 100MB, push to a HF Dataset repo: `<you>/cook-with-me-recipes`.
+4. Write `src/data/recipe_index.py`:
+   - `class RecipeIndex` with `.search(ingredients: list[str], top_k=5) -> list[RecipeRow]`.
+   - Build a query string from ingredients, embed it, cosine-similarity against the cached embeddings, return top-k.
+**Deliverable:** `python -c "from src.data.recipe_index import RecipeIndex; r=RecipeIndex(); print(r.search(['chicken','onion','tomato']))"` prints 5 sensible recipes.
+**Verify:** at least 3 of the top-5 results contain ≥2 of the input ingredients.
+---
+### Phase 2 — Day 2: Vision agent (Mise en Place) — MiniCPM-V-4.6 via llama.cpp
+**Goal:** given a fridge photo, return a clean list of English ingredient names.
+**Background:** llama.cpp supports multimodal models through a vision projector (`mmproj-*.gguf`) plus the language model GGUF. MiniCPM-V family ships both files on the Hub.
+**Tasks**
+1. Find the GGUF release of MiniCPM-V-4.6. Search HF for `MiniCPM-V-4_6-gguf` or `openbmb/MiniCPM-V-4_6-gguf`. You need **two** files:
+   - `Model-Q4_K_M.gguf` (or similar quant)
+   - `mmproj-model-f16.gguf` (the vision projector)
+2. Write `src/models/loader.py`:
+   ```python
+   from huggingface_hub import hf_hub_download
+   from llama_cpp import Llama
+   from llama_cpp.llama_chat_format import MiniCPMv26ChatHandler  # or matching handler
+   _vision = None
+   def get_vision_model():
+       global _vision
+       if _vision is None:
+           model_path = hf_hub_download(
+               repo_id="openbmb/MiniCPM-V-4_6-gguf",  # confirm exact repo
+               filename="Model-Q4_K_M.gguf",
+           )
+           mmproj_path = hf_hub_download(
+               repo_id="openbmb/MiniCPM-V-4_6-gguf",
+               filename="mmproj-model-f16.gguf",
+           )
+           handler = MiniCPMv26ChatHandler(clip_model_path=mmproj_path)
+           _vision = Llama(
+               model_path=model_path,
+               chat_handler=handler,
+               n_ctx=4096,
+               n_threads=4,
+               verbose=False,
+           )
+       return _vision
+   ```
+3. Write `src/agents/mise_en_place.py`:
+   ```python
+   import base64, io, json
+   from PIL import Image
+   from src.models.loader import get_vision_model
+   PROMPT = (
+     "You are an ingredient detector. Look at the fridge/pantry photo and "
+     "list every edible ingredient you can identify. Return strict JSON: "
+     '{"ingredients": ["chicken", "onion", "tomato", ...]} '
+     "Lowercase, English, no brand names, no containers."
+   )
+   def _img_to_data_url(img: Image.Image) -> str:
+       buf = io.BytesIO(); img.save(buf, "JPEG", quality=85)
+       b64 = base64.b64encode(buf.getvalue()).decode()
+       return f"data:image/jpeg;base64,{b64}"
+   def identify_ingredients(image: Image.Image) -> list[str]:
+       llm = get_vision_model()
+       out = llm.create_chat_completion(messages=[
+           {"role": "user", "content": [
+               {"type": "image_url", "image_url": {"url": _img_to_data_url(image)}},
+               {"type": "text", "text": PROMPT},
+           ]}
+       ], temperature=0.2, response_format={"type": "json_object"})
+       data = json.loads(out["choices"][0]["message"]["content"])
+       return [s.lower().strip() for s in data["ingredients"]]
+   ```
+4. Test locally with 5 sample fridge photos.
+**Deliverable:** the function returns a non-empty English list with ≥80% precision on a clean fridge photo.
+**Verify:** stash these 5 results in `tests/vision_smoke.json` for regression checks.
+---
+### Phase 3 — Day 3: Recipe Planner — MiniCPM-V-4 via llama.cpp + retrieval
+**Goal:** given a list of ingredients (and optionally a chosen dish), return a fully structured `Recipe` JSON including steps, durations, visual descriptions, and nutritional values.
+**Tasks**
+1. Find or convert MiniCPM-V-4 to GGUF. Likely repo: `openbmb/MiniCPM-V-4-gguf` or community quants. Pick `Q4_K_M`.
+2. Add to `src/models/loader.py` a `get_planner_model()` (same pattern as vision but without `chat_handler`).
+3. Write `src/agents/recipe_planner.py`:
+   - **Step A — propose:** call planner with `Tengo: [ingredients]. Propose 3 dish options that fit. Reply JSON.`
+   - **Step B — retrieve:** for the chosen dish name, call `RecipeIndex.search(...)` and pick the closest match. Use it as a *grounded reference*.
+   - **Step C — restructure:** prompt the planner with both the user's available ingredients and the retrieved reference recipe, asking it to output the canonical `Recipe` JSON schema below. The retrieval grounds the model and prevents hallucinated steps.
+   - **Step D — nutrition:** from the recipe ingredients, compute approximate nutritional values per serving. See Phase 3.5.
+4. Define the canonical schema in `src/pipeline.py` using Pydantic:
+   ```python
+   from pydantic import BaseModel
+   from typing import Optional
+   class Step(BaseModel):
+       n: int
+       instruction: str       # English, imperative
+       duration: str          # "4 minutes"
+       visual: str            # English visual description for FLUX prompt
+       tip: Optional[str] = None
+   class Nutrition(BaseModel):
+       calories: int          # per serving
+       protein_g: float
+       carbs_g: float
+       fat_g: float
+       fiber_g: float
+   class Recipe(BaseModel):
+       name: str
+       cuisine: str
+       servings: int
+       total_time_minutes: int
+       options: list[dict]    # only populated on "propose" call
+       ingredients_have: list[str]
+       ingredients_missing: list[str]
+       substitutes: dict[str, list[str]]
+       steps: list[Step]
+       final_dish_visual: str
+       nutrition_per_serving: Nutrition
+   ```
+5. Write the system prompt (`src/prompts/planner_system.txt`):
+   - Persona: international chef
+   - Hard rule: output JSON only, matching schema
+   - Hard rule: prefer dishes feasible with available ingredients
+   - Hard rule: 5–7 steps, each ≤ 25 words, each with a concrete `visual` field for image generation
+   - Hard rule: include `nutrition_per_serving` (model is allowed to estimate; you'll override with `data/nutrition.py` for accuracy)
+6. Use `response_format={"type": "json_object"}` in the chat completion call. Set `temperature=0.7, top_p=0.95, enable_thinking=True` for the propose step (creative); `temperature=0.4` for the structured-output step (deterministic).
+**Deliverable:** for `["chicken","onion","tomato","tortilla","cheese"]` and chosen dish "chicken tinga", the function returns a valid `Recipe` Pydantic object with 5–7 steps.
+**Verify:** the JSON parses, each step has all required fields, and total inference time on Space CPU < 60 seconds.
+---
+### Phase 3.5 — Day 3 (afternoon): Nutritional values
+**Goal:** the recipe ends with reliable per-serving nutrition (not hallucinated by the LLM).
+**Approach:** small, embedded reference table beats LLM math.
+**Tasks**
+1. Bundle `data/nutrition_table.csv` — a 200-row CSV mapping common English ingredient names to per-100g macros (kcal, protein, carbs, fat, fiber). Source: USDA FoodData Central CSV download (free, public domain). Trim columns; commit to repo.
+2. Write `src/data/nutrition.py`:
+   - `parse_quantity(line: str) -> (grams, ingredient_name)` — handle "2 cups flour", "200 g chicken", "1 tbsp olive oil". Use a small regex + a unit-to-grams table (cup=240, tbsp=15, tsp=5, oz=28.35).
+   - `compute_nutrition(ingredient_lines: list[str], servings: int) -> Nutrition` — sum per-100g values weighted by grams, divide by servings.
+   - If a line cannot be parsed, skip it and log; don't crash.
+3. After the planner returns a recipe, **overwrite** `recipe.nutrition_per_serving` with the computed value. Keep the LLM's value only as a fallback when the parser yields zero.
+**Deliverable:** for a known recipe (e.g., spaghetti with tomato sauce, 4 servings), computed calories per serving is within ±25% of online references.
+---
+### Phase 4 — Day 4: Step Illustrator — FLUX.2 Klein 9B
+**Goal:** generate an appetizing image for the final dish + one image per step.
+**Constraint:** FLUX.2 Klein on CPU is impractical; on a free Space CPU it would take ~10 minutes per image. Two paths:
+- **Path A (recommended for the hackathon):** upgrade the Space to a GPU instance (T4 or A10G — paid, but $20 HF credits cover it for a week of development). Code stays unchanged.
+- **Path B (fallback):** run FLUX in `enable_model_cpu_offload()` mode with `num_inference_steps=4` and accept ~3 min/image — only feasible for pre-rendered demo recipes, not live runs.
+**Tasks**
+1. Add to `requirements.txt`:
+   ```text
+   diffusers>=0.31
+   transformers>=4.45
+   accelerate
+   torch
+   safetensors
+   ```
+2. Write `src/models/illustrator.py`:
+   ```python
+   import torch
+   from diffusers import Flux2KleinPipeline
+   _pipe = None
+   def get_flux():
+       global _pipe
+       if _pipe is None:
+           dtype = torch.bfloat16
+           _pipe = Flux2KleinPipeline.from_pretrained(
+               "black-forest-labs/FLUX.2-klein-9B",
+               torch_dtype=dtype,
+           )
+           _pipe.enable_model_cpu_offload()
+       return _pipe
+   def render(prompt: str, seed: int = 0) -> "PIL.Image.Image":
+       pipe = get_flux()
+       device = "cuda" if torch.cuda.is_available() else "cpu"
+       img = pipe(
+           prompt=prompt,
+           height=1024, width=1024,
+           guidance_scale=1.0,
+           num_inference_steps=4,
+           generator=torch.Generator(device=device).manual_seed(seed),
+       ).images[0]
+       return img
+   ```
+3. Write `src/agents/step_illustrator.py`:
+   - For each `Step.visual`, build a prompt like:
+     > `f"Top-down photo of a kitchen pan or plate showing {visual}. {cuisine} home cooking, warm natural lighting, recipe magazine style, photorealistic, appetizing."`
+   - Generate the **final dish image first**, then the per-step images, all in **one Python loop** (no parallelism — FLUX holds the GPU).
+   - Cache results on disk keyed by `hash(prompt)` to avoid re-renders on re-runs.
+   - Emit Gradio progress updates so the UI doesn't appear frozen.
+4. **Critical tuning:** keep `num_inference_steps=4` (Klein is distilled). Higher counts blow latency and offer minimal quality gain at this scale.
+**Deliverable:** for a 5-step recipe, all 6 images (final + 5 steps) render in:
+- < 1 minute on T4 GPU Space
+- < 8 minutes on CPU offload (acceptable only for pre-cached demos)
+**Verify:** show the 6 images to an unprompted human; ≥4 should be described as "appetizing".
+---
+### Phase 5 — Day 5: Narrator — VoxCPM2
+**Goal:** every step's instruction is rendered to an MP3 in a warm, clear English voice.
+**Tasks**
+1. Confirm the exact VoxCPM2 repo name on HF (`openbmb/VoxCPM2` or similar). Read its README for the inference snippet — TTS APIs vary widely between models.
+2. Add to `requirements.txt`: `soundfile`, `torchaudio`, `numpy`. If VoxCPM2 ships GGUF, use it via `llama-cpp-python` audio extension (if available); otherwise load via `transformers` directly.
+3. Write `src/models/narrator.py`:
+   ```python
+   _tts = None
+   def get_tts():
+       global _tts
+       if _tts is None:
+           # placeholder — replace with the exact VoxCPM2 loading code from its README
+           from transformers import AutoModel, AutoProcessor
+           _tts = ... # load on CPU; VoxCPM2 is small (~1B)
+       return _tts
+   def synthesize(text: str, voice: str = "warm_female_en") -> bytes:
+       """Returns MP3 bytes."""
+       tts = get_tts()
+       wav = tts.generate(text, voice=voice)  # API depends on VoxCPM2
+       # encode wav -> mp3 with soundfile + ffmpeg-python or pydub
+       return mp3_bytes
+   ```
+4. Write `src/agents/narrator.py`:
+   - For each step, synthesize `step.instruction`. If `step.tip` is set, synthesize a separate "tip" clip.
+   - Save MP3 files in a per-recipe temp directory; return file paths to Gradio.
+5. Pre-render all step audio when the recipe is finalized — never stream per-step in the demo (too much UI lag).
+**Deliverable:** clicking "Play" on step 1 in the UI plays clear English narration.
+**Verify:** on a 5-step recipe, total TTS rendering time < 30 seconds on CPU.
+---
+### Phase 6 — Day 6: Gradio UI (Off-Brand)
+**Goal:** the Space looks like a recipe magazine, not stock Gradio.
+**Tasks**
+1. Write `src/ui/theme.py`:
+   ```python
+   import gradio as gr
+   theme = gr.themes.Soft(
+       primary_hue="orange",
+       neutral_hue="stone",
+       font=[gr.themes.GoogleFont("Inter"), "sans-serif"],
+       font_mono=[gr.themes.GoogleFont("JetBrains Mono"), "monospace"],
+   )
+   CSS = """
+   .gradio-container { background: #f5ecd9 !important; }
+   .recipe-hero { background:#fffbf0; border-radius:14px; padding:28px; }
+   .recipe-hero h1 { font-family:'Lora',serif!important; font-size:36px!important; color:#6b4a2a!important; }
+   .step-card { background:#fffbf0; border-left:4px solid #a85c2a; border-radius:8px; padding:18px 22px; margin:12px 0; }
+   .nutri-grid { display:grid; grid-template-columns:repeat(5,1fr); gap:12px; margin-top:24px; }
+   .nutri-cell { background:#fffbf0; border:1px solid #d8c9ad; border-radius:10px; padding:12px; text-align:center; }
+   """
+   ```
+2. Write `app.py` with three tabs:
+   - **Tab 1 — Cook**: fridge photo input → ingredient chips → 3 dish options → selected recipe card with hero image, steps (image + text + audio play button each), nutrition grid at the bottom.
+   - **Tab 2 — Check Progress**: upload a progress photo + select active step → validator returns badge (`go/wait/fix`) + tip + audio.
+   - **Tab 3 — About / Tech**: README-style explanation, badges, model list.
+3. Use `gr.Blocks` with `gr.State` to hold the current `Recipe` Pydantic object across UI events. Serialize to/from `dict` since Pydantic objects don't survive Gradio state by default — wrap in `state.value = recipe.model_dump()`.
+4. Wire callbacks:
+   - `btn_propose.click(fn=on_propose, inputs=[fridge_photo], outputs=[ingredient_chips, dish_options, state])`
+   - `dish_options.select(fn=on_pick_dish, inputs=[state, picked_dish], outputs=[recipe_card, hero_img, steps_column, nutrition_grid, state])`
+   - `progress_image.upload(fn=on_validate, inputs=[state, current_step_idx, progress_image], outputs=[verdict_md, tip_audio])`
+**Deliverable:** end-to-end run from a sample fridge photo to a fully rendered recipe card with audio and nutrition. No Gradio default look anywhere.
+---
+### Phase 7 — Day 7: Progress Validator (closed loop)
+**Goal:** user uploads a progress photo, app says "go / wait / fix" with a voiced tip.
+**Tasks**
+1. Write `src/agents/progress_validator.py`:
+   ```python
+   PROMPT = """Compare these two cooking photos.
+   Photo 1 (target): how it should look after the step "{instruction}".
+   Photo 2 (user's pan/plate): the user's current progress.
+   Reply strict JSON: {"verdict": "go|wait|fix", "feedback": "...", "tip": "..."}
+   - "go": looks right, move to next step
+   - "wait": needs more time, do not change anything yet
+   - "fix": something is off; suggest a concrete adjustment in one sentence
+   """
+   def validate(target_img, user_img, step_instruction): ...
+   ```
+2. Use the same vision model singleton as Phase 2 — both calls share weights.
+3. Render the verdict as a colored badge (green/amber/red) and play the tip via VoxCPM2.
+**Deliverable:** running the validator on 5 real progress photos returns the correct verdict on ≥3.
+---
+### Phase 8 — Day 8: Fine-tune the Planner on the Kaggle dataset (Well-Tuned badge)
+> **Important caveat:** The user instruction says "for now keep inference on llama.cpp inside HF Space, future migration to Modal." Fine-tuning still **requires GPU**, so training itself happens on Modal (one-shot, offline) or on a rented Colab/Lambda GPU. Inference of the resulting model stays on llama.cpp inside the Space (as GGUF). This does **not** violate the runtime constraint — only the build pipeline touches a GPU.
+**Goal:** publish a fine-tuned Planner GGUF to the Hub and load it from the Space.
+**Tasks**
+1. **Build SFT dataset** (`scripts/build_sft_dataset.py`):
+   - Load Kaggle `better-recipes` dataset.
+   - For each recipe, build a `(prompt, completion)` pair where `prompt` is `"Available ingredients: X, Y, Z. Propose recipe."` and `completion` is the full canonical `Recipe` JSON.
+   - Generate ~1000 pairs, push to `<you>/cook-with-me-sft` HF Dataset.
+2. **LoRA training** (`scripts/train_planner.py` — to be run on a GPU machine, not the Space):
+   ```python
+   # peft + trl SFTTrainer, base = openbmb/MiniCPM-V-4
+   # r=16, alpha=32, lr=2e-4, epochs=2, batch=4
+   # push_to_hub=True, hub_model_id="<you>/cook-with-me-planner-4b"
+   ```
+3. **Convert to GGUF** (Day 8 evening):
+   - Use `llama.cpp/convert_hf_to_gguf.py` then `quantize` to `Q4_K_M`.
+   - Push GGUF to `<you>/cook-with-me-planner-4b-gguf`.
+4. Update `src/models/loader.py` to point at your GGUF instead of the base model.
+**Deliverable:** the Space loads your fine-tuned Planner GGUF and produces JSON recipes that are noticeably better-formatted than the base model on a held-out test set.
+---
+### Phase 9 — Day 9: End-to-end test, performance pass, pre-warm cache
+**Goal:** the Space loads in <60s and a full recipe (text + 5 images + 5 audios + nutrition) renders in <2 minutes on the chosen hardware.
+**Tasks**
+1. Write `scripts/smoke_test.py` that runs the full pipeline on 3 sample fridge photos and asserts:
+   - Each ingredient list is non-empty
+   - Each recipe has 5–7 steps
+   - Each step has a non-empty image and audio path
+   - Nutrition has all 5 macros set
+2. Implement **on-disk caching** for FLUX outputs (key = SHA256 of prompt) so re-runs of the same recipe are instant. Save to `~/.cache/cook-with-me/flux/`.
+3. Pre-render and commit **3 fully-prepared demo recipes** (chicken tinga, pasta carbonara, chicken tikka) so judges see results in <5s on first click.
+4. Add error handling at every UI boundary: a model failure should display a friendly message, not a stack trace.
+5. Add a "Loading models..." progress bar on first request — first cold start can take 90s.
+**Deliverable:** smoke test passes on the live Space.
+---
+### Phase 10 — Day 10: README, demo video, social post, submit
+**Tasks**
+1. Write `README.md` with the required HF Space frontmatter:
+   ```yaml
+   ---
+   title: Cook With Me
+   emoji: 🍲
+   colorFrom: orange
+   colorTo: yellow
+   sdk: gradio
+   sdk_version: 4.44.0
+   app_file: app.py
+   pinned: false
+   license: apache-2.0
+   ---
+   ```
+   Followed by:
+   - One-paragraph pitch
+   - 60-second demo video embed
+   - Architecture diagram (export from `arquitectura.html` as PNG)
+   - Section: "How closed-loop visual cooking guidance works"
+   - Models used (with HF links + total parameter count)
+   - Badges declared
+   - Build / run instructions
+2. Record a 60–90 second demo video: real person cooks a recipe end-to-end with the app guiding via voice, ending with the cooked plate on camera.
+3. Write the Field Notes blog post: one of the engineering surprises (e.g., "FLUX.2 step images at 4 steps look better than 8 — here's why" or "Closed-loop validation needs the same vision model on both sides").
+4. Social post on X / LinkedIn with the demo video.
+5. Submit on the hackathon platform.
+---
+## 4. Tools usage matrix (when to reach for what)
+| Phase | Primary tools | Why |
+|---|---|---|
+| 0 — setup | HF CLI, Kaggle CLI, OpenAI Codex CLI | one-shot config |
+| 1 — data | `kagglehub`, `pandas`, `sentence-transformers` | offline dataset prep |
+| 2 — vision | `llama-cpp-python` + `MiniCPMv26ChatHandler` | runs inside Space, badge: Llama Champion |
+| 3 — planner | `llama-cpp-python` + retrieval over local parquet | grounded JSON output |
+| 3.5 — nutrition | local CSV + regex parser | reliable, no LLM math |
+| 4 — illustrator | `diffusers` + `Flux2KleinPipeline` | sponsor model showcase |
+| 5 — narrator | VoxCPM2 via `transformers` (or its native API) | local TTS |
+| 6 — UI | `gradio` + custom CSS theme | Off-Brand badge |
+| 7 — validator | same vision singleton as phase 2 | closed-loop innovation, Best Agent |
+| 8 — fine-tune | `peft`, `trl`, `llama.cpp` convert/quantize, on a GPU machine | Well-Tuned badge |
+| 9 — test/cache | `pytest`, `hashlib`, on-disk FLUX cache | demo reliability |
+| 10 — submit | HF Spaces, video tool, social | shipping |
+---
+## 5. Performance budget on the HF Space
+| Operation | Target latency | Hardware needed |
+|---|---|---|
+| Vision: ingredient ID | < 8 s | CPU 4-thread |
+| Planner: propose 3 dishes | < 12 s | CPU 4-thread |
+| Planner: build full recipe JSON | < 20 s | CPU 4-thread |
+| Nutrition computation | < 0.1 s | CPU |
+| FLUX: 1 image (4 steps) | < 12 s on T4 / < 90 s on CPU offload | GPU strongly recommended |
+| FLUX: 6 images (final + 5 steps) | < 80 s on T4 | GPU |
+| VoxCPM2: 1 step narration | < 5 s | CPU |
+| Validator: 1 progress check | < 8 s | CPU |
+| **Full recipe end-to-end** | **< 2 min on T4 Space** | — |
+**Hardware decision:** rent a T4 Space (~$0.40/hr) for the demo week. The $20 HF credits cover ~50 hours.
+---
+## 6. Risks and mitigations (delta from `estrategia.md`)
+| Risk | Mitigation |
+|---|---|
+| MiniCPM-V-4 has no public GGUF | Convert yourself with `llama.cpp/convert_hf_to_gguf.py`. Allow a half-day buffer in Phase 2. |
+| llama-cpp-python's MiniCPM-V chat handler version mismatch | Pin `llama-cpp-python==0.3.2` minimum; test the handler import on Day 2. If it fails, fall back to MiniCPM-V-2.6 GGUF (well-supported) for vision and document the swap. |
+| FLUX.2 Klein 9B too slow on free CPU Space | Upgrade to a paid GPU Space (~$10 for the demo week). Document this in the README so judges expect it. |
+| VoxCPM2 docs sparse | Drop to Kokoro-82M or Piper TTS as a backup. Lose the OpenBMB voice angle but keep the audio. |
+| Kaggle dataset has format quirks (HTML in instructions, missing fields) | The Phase 1 normalization step handles this; budget 2 hours. |
+| Nutrition CSV missing exotic ingredients | Skip-and-log strategy already designed; demo-day recipes use common ingredients only. |
+| Total params >32B if VoxCPM2 turns out to be 7B | Check size in Phase 0; if too large, drop to a smaller TTS. |
+---
+## 7. "Day-1 hello world" checklist
+Before writing any agent code, get this minimal end-to-end loop working — it proves your stack:
+1. ☐ Empty Gradio Space deployed, shows "Hello"
+2. ☐ `huggingface-cli login` works locally
+3. ☐ `kaggle datasets download thedevastator/better-recipes-for-a-better-life` succeeds
+4. ☐ `from llama_cpp import Llama` runs in your venv
+5. ☐ Download one tiny GGUF (e.g., TinyLlama Q4) and call it from a Gradio textbox round-trip
+6. ☐ Push the round-trip to the Space; confirm it answers in the cloud
+**Only after all 6 are checked, start Phase 1.**
+---
+## 8. Where this plan differs from `estrategia.md` (deltas to communicate)
+| Topic | `estrategia.md` (Spanish, Mexican-cuisine focus) | This document (current requirements) |
+|---|---|---|
+| Language | Spanish-first | **English only** |
+| Cuisine | Mexican | **International** (Kaggle dataset) |
+| Voice models | OpenBMB voice + Cohere Labs | **VoxCPM2** only (single voice) |
+| Vision model | MiniCPM-V 2.6 / 4 | **MiniCPM-V-4.6** |
+| Reasoning model | MiniCPM-4 4B | **MiniCPM-V-4** |
+| FLUX runtime | Modal endpoint | **Inside Space (llama.cpp principle)**; Modal kept as a future migration target only |
+| External APIs at runtime | Allowed (Modal, OpenAI optional) | **None** — full local inference inside Space |
+| Nutritional info | Not specified | **Required** at end of recipe |
+| Fine-tune dataset | 200 synthetic Mexican recipes | **Kaggle better-recipes (international)** |
+If anything in `plan.md` or `estrategia.md` conflicts with this document, **this document wins** — it reflects the latest user requirements.
+---
+## 9. Definition of done
+The implementation is complete when **all** of these are true:
+- [ ] Public HF Space `https://huggingface.co/spaces/<you>/cook-with-me` loads
+- [ ] App is fully in English
+- [ ] Fridge photo → ingredient list → 3 dish options → full recipe with images, audio, and nutrition works end-to-end
+- [ ] Progress validator returns sensible verdicts on 3+ test photos
+- [ ] All inference (vision, planner, TTS) runs through llama.cpp / local diffusers — **no external API calls at runtime**
+- [ ] Total parameters declared in README ≤ 32B
+- [ ] Fine-tuned Planner GGUF published to HF Hub (Well-Tuned badge)
+- [ ] Demo video (60–90s) recorded with a real person cooking
+- [ ] Field Notes blog post published
+- [ ] Submitted on the hackathon platform before deadline