INTELLIGENCE.md

Build Small Hackathon — Full Intelligence Dump

Compiled 2026-06-08

New Resources (not in original HACKATHON_README.md)

gr.Server — True Custom Frontends (Off-Brand Badge Key)

• Blog: https://huggingface.co/blog/introducing-gradio-server
• Docs: https://www.gradio.app/main/guides/server-mode
• gradio.Server inherits from FastAPI — you can serve ANY HTML/CSS/JS frontend while keeping Gradio's queuing, concurrency, SSE streaming, and ZeroGPU support.
• Frontend connects via @gradio/client JS library using client.predict()
• This is the REAL Off-Brand path — not just custom CSS on gr.Blocks
• Action: Upgrade FocusFriend to gr.Server with a fully custom vanilla HTML/JS frontend

ML Intern — Autonomous ML Agent

• Repo: https://github.com/huggingface/ml-intern
• Autonomous agent that researches, writes, and ships ML code using HF ecosystem
• Can be used WITH our projects for Sharing is Caring badge (agent traces)
• CLI tool: ml-intern "build a TTS app" — it researches and codes autonomously
• Action: Run ml-intern on our projects, share traces to Hub

gr.Server Code Pattern (for FocusFriend upgrade)

from gradio import Server
from fastapi.responses import HTMLResponse

app = Server()

@app.api(name="chat")
def chat_endpoint(message: str, history: list) -> dict:
    # LLM inference here
    return {"response": "...", "mood": "focus"}

@app.get("/", response_class=HTMLResponse)
async def homepage():
    with open("index.html") as f:
        return f.read()

app.launch()

Frontend (index.html) uses vanilla HTML/CSS/JS + @gradio/client library:

import { Client } from "@gradio/client";
const client = await Client.connect(window.location.origin);
const result = await client.predict("/chat", { message, history });

---

Competition Landscape — What's Already Submitted

Spaces in the Org (213 total — 6 sampled)

Space	Concept	Track
mind-of-tashi	AI consciousness sim	Thousand Token Wood
vivamais	Health/wellness AI	Backyard AI
split-brain-copilot	Dual-perspective AI assistant	Thousand Token Wood
lolaby	Lullaby/audio generation	Backyard AI
NPCverse	AI NPC dialogue system	Thousand Token Wood
museum-of-unlived-lives	Generative art/storytelling	Thousand Token Wood

Models Published (32 total)

Notable fine-tunes (proving Well-Tuned badge is achievable):

• mind-of-tashi-micro-grpo (GRPO fine-tune)
• NeuroBait
• flux-costume-booth-lora (LoRA for Flux)
• tiny-quest-radio-aya-lora (LoRA fine-tune)
• pocket-tutor-minicpmv-socratic

Datasets Published (22 total) — Sharing is Caring in action

• mind-of-tashi-runs (agent traces)
• 1000-Rooms-traces
• hackathon-advisor-codex-traces
• bedtime-story-machine-trace
• jawbreaker-scam-defense-data

Blog Posts (12 Field Notes examples)

• ClinIQ: On-device pharmacist
• Thousand Token Wood sim v2/v3
• Building Pakistan Notice Helper
• Amazing Digital Dentures
• Room360: Video to 3D reconstruction
• Mythograph Atelier
• Her · हेर
• Job Searcher
• Persona Atlas

---

Gap Analysis — What We Haven't Covered

Badges NOT Yet Targeted

Badge	Status	How to Get It
🦙 Llama Champion	❌	Need to run model through llama.cpp runtime directly (not just llama-cpp-python)
📡 Sharing is Caring	❌	Need to publish agent traces to HF Hub
🎨 Off-Brand (true)	⚠️ Partial	FocusFriend has custom CSS but should use gr.Server for full custom frontend
🎯 Well-Tuned	❌	Need to actually fine-tune and publish a model

Sponsor Prizes NOT Targeted

Sponsor	Prize	How to Target
NVIDIA	2× RTX 5080	Build with Nemotron models
Cohere	$5K cash	Use Cohere models (Command R, etc.)
Black Forest Labs	$3K cash	Use Flux image models
Modal	$20K credits	Deploy on Modal infrastructure

Special Awards NOT Targeted

Award	Prize	How to Target
Tiny Titan	$1,000	App using ≤4B param model
Bonus Quest Champion	$2,000	Most badges on single submission
Best Demo	$1,500	Best video + social post package

---

Third Project Ideas — Fill the Gaps

Option A: "Tiny Bard" — Micro-Adventure Generator

• Track: Thousand Token Wood
• Model: Qwen 2.5 1.5B or Phi-4-mini (3.8B) — Tiny Titan eligible
• Runtime: llama.cpp (Llama Champion badge)
• Concept: Tiny LLM generates interactive text adventures. Each playthrough is 5 minutes. The AI creates rooms, NPCs, items, and branching narratives on the fly.
• Frontend: gr.Server with custom pixel-art/terminal UI
• Badges: 🦙 Llama Champion · 🏆 Tiny Titan · 🎨 Off-Brand · 🔌 Off the Grid · 📓 Field Notes
• Prize targets: Tiny Titan ($1K) · Thousand Token Wood ($4K-$1K) · Bonus Quest Champion potential

Option B: "CritterCalm Mini" — Tiny Titan Edition

• Track: Backyard AI
• Model: Replace Dolphin-X1-8B with Qwen 2.5 1.5B — Tiny Titan eligible
• Concept: Slim version of CritterCalm that runs on ANY hardware. Just the calming templates + Kokoro TTS.
• Badges: 🏆 Tiny Titan · 🔌 Off the Grid
• Effort: Low (mostly repackaging existing work)

Option C: "DreamSnap" — Surreal Pet Portrait Generator

• Track: Thousand Token Wood
• Model: Black Forest Labs Flux (small variant) + small LLM
• Concept: Describe your pet → generate surreal/delightful pet portraits. AI co-creates the art.
• Badges: 🎨 Off-Brand · 📓 Field Notes
• Prize targets: Black Forest Labs sponsor prize ($3K) · Thousand Token Wood

Option D: "TraceShare" — Agent Trace Publisher

• Concept: Utility tool that helps ALL hackathon participants share agent traces easily
• Use ML Intern to generate traces, then publish to HF Hub
• Badges: 📡 Sharing is Caring (applied to multiple projects)
• Effort: Very low — just run ml-intern, capture traces, publish

---

Recommended Action Plan (June 8-15)

Priority 1: Upgrade Existing Projects

1. FocusFriend → gr.Server — Replace gr.Blocks with fully custom HTML/JS frontend
   • True Off-Brand badge qualification
   • Keep all existing Pip personality + mode logic
2. CritterCalm → Add llama.cpp runtime — Get Llama Champion badge
3. Both → Share agent traces — Run ml-intern, publish traces to Hub

Priority 2: Quick Win — Tiny Titan

4. Build "CritterCalm Mini" — Slim version with ≤4B params
   • Same app structure, smaller model
   • Gets Tiny Titan award eligibility

Priority 3: Third Project (if time permits)

5. Build "Tiny Bard" — Micro-adventure generator
   • Hits 5 badges at once (Llama Champion, Tiny Titan, Off-Brand, Off the Grid, Field Notes)
   • Strong Bonus Quest Champion candidate

---

Updated Badge Strategy Per Project

CritterCalm

Badge	Status	Action
🔌 Off the Grid	✅ Built local	Verify no API calls
🎯 Well-Tuned	⬜	Fine-tune OmniVoice on pet-directed speech
📓 Field Notes	✅ Draft written	Polish and publish
📡 Sharing is Caring	⬜	Run with ml-intern, publish traces
🦙 Llama Champion	⬜	Add llama.cpp server mode option

FocusFriend

Badge	Status	Action
🎨 Off-Brand	⚠️ Partial	Upgrade to gr.Server custom frontend
🔌 Off the Grid	✅ Built local	Verify no API calls
📓 Field Notes	✅ Draft written	Polish and publish
📡 Sharing is Caring	⬜	Run with ml-intern, publish traces
🦙 Llama Champion	⚠️ Partial	Already uses llama-cpp-python — add llama.cpp server flag

Tiny Bard (New)

Badge	Status
🦙 Llama Champion	✅ llama.cpp native
🏆 Tiny Titan	✅ ≤4B model
🎨 Off-Brand	✅ gr.Server custom UI
🔌 Off the Grid	✅ Fully local
📓 Field Notes	To write

---

Links Repository

Hackathon Official

• Org: https://huggingface.co/build-small-hackathon
• Discord: https://discord.gg/YHECTft87Z
• README Space: https://huggingface.co/spaces/build-small-hackathon/README

Tools & Kits

• ML Intern: https://github.com/huggingface/ml-intern
• Gradio Quickstart: https://www.gradio.app/guides/quickstart
• gr.Server Blog: https://huggingface.co/blog/introducing-gradio-server
• gr.Server Docs: https://www.gradio.app/main/guides/server-mode
• llama.cpp: https://github.com/ggml-org/llama.cpp

Our Repos

• Monorepo: (this)
• CritterCalm: https://github.com/nbiish/crittercalm
• FocusFriend: https://github.com/nbiish/focusfriend

Models We're Using

• OmniVoice: https://github.com/k2-fsa/OmniVoice
• Dolphin-X1-8B: https://huggingface.co/dphn/Dolphin-X1-8B-GGUF
• Gemma 4 12B: https://huggingface.co/unsloth/gemma-4-12b-it-GGUF
• Kokoro TTS: https://github.com/thewh1teagle/kokoro-onnx