Back to AI & ML The Lab

Models and agents, live in your browser

Working demos, not screenshots. Everything on this page runs on your device — the sentiment model executes on your GPU via WebGL, and nothing you type leaves the browser.

Live Lab

Try a model in your browser

A small CNN trained on movie reviews, running entirely on your device — no server, no upload. Type a sentence and the model classifies its sentiment in milliseconds. Lazy-loads when this section enters view.

Awaiting input — try a sample or type your own.
Model
sentiment-CNN v1
Size
—
Inference
—
Runtime
TensorFlow.js · WebGL
Under the hood
  1. Tokenise — split your sentence on whitespace and look up each word in a 20k-word vocabulary.
  2. Encode — map words to integer indices, pad or truncate to length 100.
  3. Inference — feed the sequence through a small CNN (embedding → conv → global max pool → dense).
  4. Decode — output is a single value 0–1 — closer to 1 is more positive.

Model and vocabulary are fetched from the public TensorFlow.js model host on first use, then cached by your browser. No data leaves your device. The whole forward pass runs on your GPU via WebGL.

In motion

See how an agent thinks

Hand-rolled animation of a research agent — input, planning, tool use, memory, synthesis, output. Auto-plays once when it scrolls into view.

  1. User input

    "Analyse this quarterly report and summarise the key findings."

  2. Planning

    Decompose into sub-tasks:

    • Extract text
    • Identify sections
    • Analyse financials
    • Detect trends
  3. Tool selection

    Choose tools for each sub-task:

    • document.parse
    • nlp.classify
    • chart.render
  4. Execution loop

    Call each tool, capture output:

    • document.parse(report.pdf) → 28 pages, 14 tables
    • nlp.classify(sections) → 4 categories
    • chart.render(trends) → 3 charts
  5. Memory

    Findings stored in working memory:

    Revenue +12% YoY Cost ratio improved EMEA outperformed
  6. Synthesis

    Combine findings into a coherent narrative; check against the original prompt.

  7. Output

    "Q3 revenue grew 12% YoY driven by EMEA. Margin improved 180bps as cost ratios fell. Forecast a softer Q4 on currency drag."