Spaces:

GoktugD
/

DUSUNEN-AI-Lab

Running

File size: 22,759 Bytes

<!doctype html>
<html lang="en">
<head>
  <meta charset="utf-8">
  <meta name="viewport" content="width=device-width, initial-scale=1">
  <meta name="theme-color" content="#080b14">
  <meta name="description" content="Göktuğ Düşünen — open Turkish AI, retrieval and security systems with reproducible evidence.">
  <title>Göktuğ Düşünen · Turkish AI Systems</title>
  <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=DM+Mono:wght@400;500&family=Manrope:wght@400;500;600;700;800&display=swap" rel="stylesheet">
  <style>
    :root {
      --bg: #080b14;
      --surface: rgba(18, 24, 42, .72);
      --surface-strong: #12182a;
      --line: rgba(255, 255, 255, .09);
      --text: #f5f7ff;
      --muted: #9ca7bd;
      --violet: #a78bfa;
      --cyan: #58d5e8;
      --lime: #b8f28d;
      --max: 1160px;
    }

    * { box-sizing: border-box; }
    html { scroll-behavior: smooth; }
    body {
      margin: 0;
      color: var(--text);
      background:
        radial-gradient(circle at 10% -10%, rgba(99, 102, 241, .25), transparent 34rem),
        radial-gradient(circle at 95% 20%, rgba(34, 211, 238, .13), transparent 30rem),
        var(--bg);
      font-family: "Manrope", sans-serif;
      min-height: 100vh;
      overflow-x: hidden;
    }

    body::before {
      content: "";
      position: fixed;
      inset: 0;
      pointer-events: none;
      opacity: .19;
      background-image:
        linear-gradient(var(--line) 1px, transparent 1px),
        linear-gradient(90deg, var(--line) 1px, transparent 1px);
      background-size: 52px 52px;
      mask-image: linear-gradient(to bottom, black, transparent 75%);
    }

    a { color: inherit; }
    .shell { width: min(var(--max), calc(100% - 40px)); margin: 0 auto; }

    nav {
      height: 84px;
      display: flex;
      align-items: center;
      justify-content: space-between;
      border-bottom: 1px solid var(--line);
    }

    .brand { display: flex; align-items: center; gap: 11px; text-decoration: none; font-weight: 800; }
    .brand-mark {
      width: 32px; height: 32px; display: grid; place-items: center;
      border: 1px solid rgba(167, 139, 250, .5); border-radius: 10px;
      background: rgba(167, 139, 250, .1); color: var(--violet);
      font-family: "DM Mono"; font-size: 13px;
    }
    .nav-links { display: flex; align-items: center; gap: 26px; }
    .nav-links a { text-decoration: none; color: var(--muted); font-size: 14px; font-weight: 600; }
    .nav-links a:hover { color: var(--text); }

    .hero {
      min-height: 610px;
      display: grid;
      grid-template-columns: 1.35fr .65fr;
      align-items: center;
      gap: 80px;
      padding: 90px 0 70px;
    }
    .hero > * { min-width: 0; }

    .eyebrow, .kicker {
      display: inline-flex; align-items: center; gap: 9px;
      color: var(--cyan); font: 500 12px/1.2 "DM Mono";
      text-transform: uppercase; letter-spacing: .13em;
    }
    .eyebrow::before { content: ""; width: 7px; height: 7px; border-radius: 50%; background: var(--cyan); box-shadow: 0 0 20px var(--cyan); }
    h1 { margin: 25px 0 24px; font-size: clamp(54px, 7vw, 92px); line-height: .98; letter-spacing: -.065em; }
    .gradient { color: transparent; background: linear-gradient(100deg, #fff 10%, var(--violet) 55%, var(--cyan)); background-clip: text; }
    .lead { max-width: 700px; margin: 0; color: var(--muted); font-size: clamp(17px, 2vw, 21px); line-height: 1.75; }
    .lead strong { color: var(--text); font-weight: 600; }
    .actions { display: flex; flex-wrap: wrap; gap: 12px; margin-top: 34px; }
    .button {
      display: inline-flex; align-items: center; justify-content: center; gap: 9px;
      min-height: 48px; padding: 0 19px; border-radius: 13px;
      border: 1px solid var(--line); text-decoration: none; font-weight: 700; font-size: 14px;
      transition: transform .2s, background .2s, border-color .2s;
    }
    .button:hover { transform: translateY(-2px); border-color: rgba(255,255,255,.25); }
    .button.primary { color: #080b14; background: var(--text); }
    .button.secondary { background: rgba(255,255,255,.04); }

    .portrait-wrap { position: relative; justify-self: end; }
    .portrait-wrap::before { content: ""; position: absolute; inset: -26px; border: 1px solid var(--line); border-radius: 50%; animation: spin 18s linear infinite; }
    .portrait-wrap::after { content: "MODELS / DATA / SYSTEMS"; position: absolute; left: -44px; bottom: 8px; padding: 10px 14px; color: var(--lime); background: #101525; border: 1px solid var(--line); border-radius: 10px; font: 500 10px "DM Mono"; letter-spacing: .12em; }
    .portrait { width: min(280px, 30vw); aspect-ratio: 1; object-fit: cover; border-radius: 50%; border: 8px solid rgba(255,255,255,.05); filter: saturate(.85) contrast(1.05); }
    @keyframes spin { to { transform: rotate(360deg); } }

    .proof-strip {
      display: grid; grid-template-columns: repeat(4, 1fr); margin: 0 0 82px;
      border: 1px solid var(--line); border-radius: 18px; overflow: hidden;
      background: rgba(255,255,255,.025); backdrop-filter: blur(18px);
    }
    .proof { padding: 22px 24px; border-right: 1px solid var(--line); }
    .proof:last-child { border-right: 0; }
    .proof strong { display: block; font-size: 28px; letter-spacing: -.04em; }
    .proof span { color: var(--muted); font: 400 10px/1.4 "DM Mono"; letter-spacing: .08em; text-transform: uppercase; }

    section { padding: 82px 0; border-top: 1px solid var(--line); }
    .section-head { display: grid; grid-template-columns: .55fr 1.45fr; gap: 50px; margin-bottom: 42px; }
    h2 { margin: 0; font-size: clamp(32px, 5vw, 54px); line-height: 1.05; letter-spacing: -.045em; }
    .section-copy { margin: 0; color: var(--muted); font-size: 17px; line-height: 1.7; max-width: 700px; }

    .projects { display: grid; grid-template-columns: repeat(12, 1fr); gap: 16px; }
    .project {
      position: relative; min-height: 270px; padding: 28px; overflow: hidden;
      border: 1px solid var(--line); border-radius: 22px; background: var(--surface);
      text-decoration: none; backdrop-filter: blur(18px);
      transition: transform .25s, border-color .25s, background .25s;
    }
    .project:hover { transform: translateY(-4px); border-color: rgba(167,139,250,.4); background: rgba(25,33,56,.84); }
    .project.large { grid-column: span 7; }
    .project.small { grid-column: span 5; }
    a.project::after { content: "↗"; position: absolute; top: 25px; right: 27px; color: var(--muted); font-size: 20px; }
    .project-no { color: var(--violet); font: 500 11px "DM Mono"; letter-spacing: .13em; }
    .project h3 { margin: 64px 0 12px; max-width: 430px; font-size: 26px; letter-spacing: -.03em; }
    .project p { margin: 0; max-width: 510px; color: var(--muted); line-height: 1.65; }
    .project.featured-oku {
      min-height: 450px; background-image:
        linear-gradient(180deg, rgba(7,11,26,.04) 30%, rgba(7,11,26,.98) 76%),
        url("assets/oku-hero.png");
      background-size: cover; background-position: left center;
    }
    .project.featured-oku h3 { margin-top: 245px; }
    .project.featured-oku p { color: #c3cee4; }
    .tags { display: flex; flex-wrap: wrap; gap: 7px; margin-top: 22px; }
    .tag { padding: 6px 9px; border: 1px solid var(--line); border-radius: 7px; color: #bac4d7; font: 400 10px "DM Mono"; }

    .pipeline { display: grid; grid-template-columns: repeat(5, 1fr); gap: 10px; counter-reset: stage; }
    .stage { position: relative; min-height: 190px; padding: 24px; border: 1px solid var(--line); border-radius: 17px; background: rgba(255,255,255,.025); }
    .stage::before { counter-increment: stage; content: "0" counter(stage); display: block; color: var(--violet); font: 500 11px "DM Mono"; letter-spacing: .13em; margin-bottom: 36px; }
    .stage:not(:last-child)::after { content: "→"; position: absolute; z-index: 2; right: -18px; top: 50%; width: 26px; height: 26px; display: grid; place-items: center; border: 1px solid var(--line); border-radius: 50%; background: var(--bg); color: var(--cyan); }
    .stage strong { display: block; margin-bottom: 8px; font-size: 16px; }
    .stage p { margin: 0; color: var(--muted); font-size: 13px; line-height: 1.55; }

    .stack { display: grid; grid-template-columns: repeat(4, 1fr); gap: 10px; }
    .stack-item { padding: 22px; border: 1px solid var(--line); border-radius: 16px; background: rgba(255,255,255,.025); }
    .stack-item span { display: block; color: var(--cyan); font: 500 10px "DM Mono"; text-transform: uppercase; letter-spacing: .12em; margin-bottom: 10px; }
    .stack-item strong { font-size: 15px; }

    footer { display: flex; justify-content: space-between; align-items: center; gap: 20px; padding: 38px 0 50px; color: var(--muted); font-size: 13px; }
    footer a { color: var(--text); text-decoration: none; }

    @media (max-width: 800px) {
      .nav-links a:first-child { display: none; }
      .hero { grid-template-columns: 1fr; gap: 70px; padding-top: 70px; }
      .portrait-wrap { justify-self: center; }
      .portrait { width: 210px; }
      .section-head { grid-template-columns: 1fr; gap: 22px; }
      .project.large, .project.small { grid-column: span 12; }
      .stack { grid-template-columns: repeat(2, 1fr); }
      .proof-strip { grid-template-columns: repeat(2, 1fr); }
      .proof:nth-child(2) { border-right: 0; }
      .proof:nth-child(-n+2) { border-bottom: 1px solid var(--line); }
      .pipeline { grid-template-columns: repeat(2, 1fr); }
      .stage::after { display: none !important; }
    }

    @media (max-width: 600px) {
      .shell { width: calc(100% - 28px); max-width: var(--max); }
      nav { height: 72px; }
      .nav-links { display: none; }
      .hero { min-height: auto; padding: 62px 0; }
      .eyebrow { display: flex; max-width: 100%; font-size: 9px; letter-spacing: .08em; line-height: 1.5; overflow-wrap: anywhere; }
      h1 { font-size: 40px; overflow-wrap: anywhere; }
      .lead { overflow-wrap: anywhere; }
      .actions { display: grid; grid-template-columns: 1fr; }
      .button { width: 100%; }
      section { padding: 62px 0; }
      .project { padding: 23px; }
      .stack { grid-template-columns: 1fr; }
      .proof-strip, .pipeline { grid-template-columns: 1fr; }
      .proof { border-right: 0; border-bottom: 1px solid var(--line); }
      .proof:last-child { border-bottom: 0; }
      footer { align-items: flex-start; flex-direction: column; }
    }

    @media (prefers-reduced-motion: reduce) {
      html { scroll-behavior: auto; }
      .portrait-wrap::before { animation: none; }
      * { transition-duration: 0s !important; }
    }
  </style>
</head>
<body>
  <div class="shell">
    <nav>
      <a class="brand" href="#top"><span class="brand-mark">GD</span><span>Göktuğ Düşünen</span></a>
      <div class="nav-links">
        <a href="#work">DUSUNEN</a>
        <a href="#nanosoc">NanoSOC1</a>
        <a href="https://github.com/Goktug-Dusunen" target="_blank" rel="noreferrer">GitHub ↗</a>
      </div>
    </nav>

    <main id="top">
      <header class="hero">
        <div>
          <div class="eyebrow">Turkish AI · OCR · speech · retrieval · security</div>
          <h1>I build Turkish AI.<br><span class="gradient">End to end.</span></h1>
          <p class="lead">I'm Göktuğ. I turn raw data into <strong>trained models, reproducible benchmarks and working inference products.</strong> Every claim links back to an artifact you can inspect, run and measure.</p>
          <div class="actions">
            <a class="button primary" href="https://huggingface.co/spaces/GoktugD/DUSUNEN-Oku-Demo" target="_blank" rel="noreferrer">Try DUSUNEN Oku <span>↗</span></a>
            <a class="button secondary" href="https://huggingface.co/GoktugD/DUSUNEN-Oku-62M-v1" target="_blank" rel="noreferrer">Inspect the model <span>↗</span></a>
          </div>
        </div>
        <div class="portrait-wrap">
          <img class="portrait" src="https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/ZSbELOIjcBxAKAL92_A-z.jpeg" alt="Göktuğ Düşünen">
        </div>
      </header>

      <div class="proof-strip" aria-label="DUSUNEN project facts">
        <div class="proof"><strong>250K</strong><span>word-disjoint OCR images</span></div>
        <div class="proof"><strong>244M</strong><span>compact Turkish ASR</span></div>
        <div class="proof"><strong>100K</strong><span>curated train triplets</span></div>
        <div class="proof"><strong>5</strong><span>held-out retrieval tasks</span></div>
      </div>

      <section id="work">
        <div class="section-head">
          <div class="kicker">01 / Flagship stack</div>
          <div>
            <h2>DUSUNEN Model Lab.</h2>
            <p class="section-copy">A measured Turkish AI program across vision, speech and search: compact OCR, ASR, text restoration, retrieval, pinned benchmarks and live inference.</p>
          </div>
        </div>

        <div class="projects">
          <a class="project large featured-oku" href="https://huggingface.co/GoktugD/DUSUNEN-Oku-62M-v1" target="_blank" rel="noreferrer">
            <div class="project-no">MODEL / TURKISH OCR</div>
            <h3>DUSUNEN Oku 62M · Turkish OCR</h3>
            <p>A compact recognizer for cropped printed Turkish words: 98.39% exact match and 0.24% CER on 12,500 vocabulary-disjoint test images.</p>
            <div class="tags"><span class="tag">98.39% EXACT</span><span class="tag">0.24% CER</span><span class="tag">LIVE DEMO</span></div>
          </a>

          <a class="project small" href="https://huggingface.co/GoktugD/DUSUNEN-Dinle-244M-v1" target="_blank" rel="noreferrer">
            <div class="project-no">MODEL / TURKISH ASR</div>
            <h3>DUSUNEN Dinle 244M</h3>
            <p>Compact Turkish speech recognition with 14.41% WER on FLEURS and a measured 12.83% relative improvement over untouched Whisper-small.</p>
            <div class="tags"><span class="tag">WHISPER</span><span class="tag">14.41% WER</span><span class="tag">AUDIO DEMO</span></div>
          </a>

          <a class="project small" href="https://huggingface.co/GoktugD/DUSUNEN-Nokta-68M-v1" target="_blank" rel="noreferrer">
            <div class="project-no">MODEL / TEXT RESTORATION</div>
            <h3>DUSUNEN Nokta 68M</h3>
            <p>One pass for Turkish punctuation and true-casing, reaching 70.43% punctuation macro F1 and 94.04% joint token accuracy.</p>
            <div class="tags"><span class="tag">68M PARAMS</span><span class="tag">JOINT LABELS</span><span class="tag">ONNX</span></div>
          </a>

          <a class="project large" href="https://huggingface.co/GoktugD/DUSUNEN-Atlas-278M-v1" target="_blank" rel="noreferrer">
            <div class="project-no">MODEL / MATRYOSHKA RETRIEVER</div>
            <h3>DUSUNEN Atlas 278M</h3>
            <p>One Turkish retrieval model with six useful embedding sizes. At 128 dimensions it cuts index storage by 83.3% while reaching 94.75% held-out hard-negative accuracy.</p>
            <div class="tags"><span class="tag">768→64-D</span><span class="tag">MATRYOSHKA</span><span class="tag">BF16</span></div>
          </a>

          <a class="project small" href="https://huggingface.co/GoktugD/DUSUNEN-Pusula-118M-v1" target="_blank" rel="noreferrer">
            <div class="project-no">MODEL / COMPACT RETRIEVER</div>
            <h3>DUSUNEN Pusula 118M</h3>
            <p>A 117.7M encoder with six selectable dimensions, a 91.7% index-size reduction at 32-D, and a pinned five-task Turkish scorecard.</p>
            <div class="tags"><span class="tag">384→32-D</span><span class="tag">MATRYOSHKA</span><span class="tag">5-TASK MTEB</span></div>
          </a>

          <a class="project small" href="https://huggingface.co/GoktugD/DUSUNEN-Mercek-118M-v1" target="_blank" rel="noreferrer">
            <div class="project-no">MODEL / CROSS-ENCODER</div>
            <h3>DUSUNEN Mercek 118M</h3>
            <p>LambdaLoss learning-to-rank over 50K lists, with measured MRR@10 and nDCG@10 gains over its untouched multilingual base on 102,400 frozen pairs.</p>
            <div class="tags"><span class="tag">LAMBDA LOSS</span><span class="tag">102.4K PAIRS</span><span class="tag">HELD-OUT UPLIFT</span></div>
          </a>

          <a class="project large" href="https://huggingface.co/datasets/GoktugD/DUSUNEN-Turkish-Retrieval-Benchmark-v1" target="_blank" rel="noreferrer">
            <div class="project-no">BENCHMARK / FIVE TURKISH TASKS</div>
            <h3>One protocol. Four systems. Raw evidence.</h3>
            <p>TurHistQuad, XQuAD, WebFAQ, MKQA and Belebele through the official MTEB evaluator, with pinned revisions and explicit model prompt formats.</p>
            <div class="tags"><span class="tag">MTEB</span><span class="tag">NDCG@10</span><span class="tag">REPRODUCIBLE</span></div>
          </a>

          <a class="project small" href="https://huggingface.co/datasets/GoktugD/DUSUNEN-HardNegatives-50K-v1" target="_blank" rel="noreferrer">
            <div class="project-no">DATASET / MINED NEGATIVES</div>
            <h3>50K difficult training triples</h3>
            <p>Model-mined negatives from a 70K candidate pool, with lexical safeguards, zero fallbacks, checksums and an exact-overlap audit.</p>
            <div class="tags"><span class="tag">HNSW</span><span class="tag">PARQUET</span><span class="tag">0 EXACT OVERLAP</span></div>
          </a>

          <a class="project large" href="https://huggingface.co/spaces/GoktugD/DUSUNEN-Cited-RAG" target="_blank" rel="noreferrer">
            <div class="project-no">SPACE / CITED RETRIEVAL</div>
            <h3>Ask in Turkish. Inspect the sources.</h3>
            <p>A free browser-only retrieval demo that selects extractive evidence and links every answer fragment to its source. No hosted model API and no hidden generation step.</p>
            <div class="tags"><span class="tag">CITED RAG</span><span class="tag">TRANSFORMERS.JS</span><span class="tag">0 PAID API</span></div>
          </a>

          <a class="project small" href="https://huggingface.co/spaces/GoktugD/DUSUNEN-Vector-Lab" target="_blank" rel="noreferrer">
            <div class="project-no">SPACE / VECTOR BUDGET LAB</div>
            <h3>Watch retrieval change from 384-D to 32-D</h3>
            <p>One browser query, six live Matryoshka vector budgets, instant ranking changes and explicit index-size trade-offs.</p>
            <div class="tags"><span class="tag">384→32-D</span><span class="tag">ONNX Q8</span><span class="tag">LOCAL INFERENCE</span></div>
          </a>
        </div>
      </section>

      <section id="nanosoc">
        <div class="section-head">
          <div class="kicker">02 / Security AI</div>
          <div>
            <h2>NanoSOC1:8B.</h2>
            <p class="section-copy">A gated Foundation-Sec 8B QLoRA system for structured SOC triage, evidence correlation, MITRE ATT&amp;CK attribution and human-approved response—published with both its benchmark gains and its failure modes.</p>
          </div>
        </div>
        <div class="projects">
          <a class="project large" href="https://huggingface.co/GoktugD/nanosoc1-8b" target="_blank" rel="noreferrer">
            <div class="project-no">MODEL / SECURITY COPILOT</div>
            <h3>NanoSOC1 8B · v6 evidence stack</h3>
            <p>Two task-routed adapters, a 5,492-document security RAG corpus, model-security gates and immutable audit evidence. Built to assist analysts—not to act as an autonomous IDS or response engine.</p>
            <div class="tags"><span class="tag">FOUNDATION-SEC 8B</span><span class="tag">QLORA</span><span class="tag">HUMAN IN THE LOOP</span></div>
          </a>

          <a class="project small" href="https://huggingface.co/spaces/GoktugD/nanosoc1-security-lab" target="_blank" rel="noreferrer">
            <div class="project-no">SPACE / EVALUATION LAB</div>
            <h3>Inspect the evidence, including failures</h3>
            <p>A browser-only evaluation companion with frozen holdout metrics, sanitized triage walkthroughs, architecture boundaries and explicit false-positive limitations.</p>
            <div class="tags"><span class="tag">96.97% DNS/C2 RECALL</span><span class="tag">FROZEN HOLDOUT</span><span class="tag">0 PAID API</span></div>
          </a>
        </div>
      </section>

      <section>
        <div class="section-head">
          <div class="kicker">03 / Pipeline</div>
          <div>
            <h2>No missing middle.</h2>
            <p class="section-copy">The work connects data decisions to measured model behavior and finally to something people can use.</p>
          </div>
        </div>
        <div class="pipeline">
          <div class="stage"><strong>Curate</strong><p>Normalize, filter, deduplicate, checksum and publish the exact training split.</p></div>
          <div class="stage"><strong>Train</strong><p>Contrastive fine-tuning with memory-aware batches on a single 8 GB GPU.</p></div>
          <div class="stage"><strong>Mine</strong><p>Retrieve difficult negatives, reject risky matches and record every mining decision.</p></div>
          <div class="stage"><strong>Evaluate</strong><p>Pinned held-out data, exact vector search, strong baselines and raw JSON results.</p></div>
          <div class="stage"><strong>Ship</strong><p>Open weights, documented inference and a live semantic-search experience.</p></div>
        </div>
      </section>

      <section>
        <div class="section-head">
          <div class="kicker">04 / Engineering</div>
          <h2>Built for scrutiny, not screenshots.</h2>
        </div>
        <div class="stack">
          <div class="stack-item"><span>Provenance</span><strong>Licenses · revisions · checksums</strong></div>
          <div class="stack-item"><span>Training</span><strong>PyTorch · CUDA · BF16</strong></div>
          <div class="stack-item"><span>Evaluation</span><strong>FAISS · MRR · nDCG · recall</strong></div>
          <div class="stack-item"><span>Delivery</span><strong>HF Hub · ONNX · Static Spaces</strong></div>
        </div>
      </section>
    </main>

    <footer>
      <span>Train it. Measure it. Ship it.</span>
      <span><a href="https://huggingface.co/GoktugD" target="_blank" rel="noreferrer">Hugging Face</a> · <a href="https://github.com/Goktug-Dusunen" target="_blank" rel="noreferrer">GitHub</a> · <a href="https://huggingface.co/Werea-co" target="_blank" rel="noreferrer">Werea</a></span>
    </footer>
  </div>
</body>
</html>