PROJECT: BOB

Local-first, serverless, tool-calling AI running entirely in your browser via WebAssembly.

---

What is Bob?

Bob is a personal AI assistant that runs 100% locally in your browser. No servers, no cloud, no tracking, no API keys required after initial model download (~600MB). Built with Transformers.js and Gorilla OpenFunctions.

Key Features

• 🔒 Zero Server — Runs entirely client-side via WebAssembly
• 🛠 Tool Calling — Native function calling: get_crypto_price, multiply_numbers
• 🌐 Offline-Capable — Works after first load (model cached in IndexedDB)
• ⚡ Instant Deploy — Static files on GitHub Pages
• 🎨 Stark UI — Pure black/white high-contrast interface

---

Live Demo

https://yummyfiles.github.io/Project-Bob/

---

Architecture

┌─────────────────────────────────────────────────────────────┐
│                      BROWSER (Client)                        │
├─────────────────────────────────────────────────────────────┤
│  index.html  →  app.js  →  @xenova/transformers (WASM)      │
│                              ↓                               │
│                    ┌──────────────────┐                      │
│                    │  Gorilla-3B-4bit │  ← Cached in         │
│                    │  (ONNX quantized)│  ← IndexedDB         │
│                    └────────┬─────────┘                      │
│                             │                                │
│                    <call_tool> detection                     │
│                             ↓                                │
│              ┌──────────────┴──────────────┐                 │
│              ▼                             ▼                 │
│      get_crypto_price()            multiply_numbers()       │
│      (CoinGecko API)               (local JS math)          │
└─────────────────────────────────────────────────────────────┘

---

Quick Start

Use Online (No Setup)

Visit https://yummyfiles.github.io/Project-Bob/ — loads model on first visit.

Run Locally

git clone https://github.com/yummyfiles/Project-Bob
cd Project-Bob
npx serve  # or python -m http.server 8000
# Open http://localhost:3000

---

Model

Component	Details
Base	gorilla-llm/gorilla-openfunctions-v2 (3B params)
Format	ONNX INT4 quantized via Optimum
Runtime	Transformers.js (ONNX Runtime Web + WebAssembly)
Size	~600MB download, cached after first load
Capabilities	Function calling, chat, reasoning

Why Gorilla?

• Trained specifically for function calling (API interaction)
• Open-source, Apache 2.0 license
• Works well quantized to 4-bit
• Small enough for browser inference

---

Tool System

Bob detects <call_tool> tags in model output and executes matching JavaScript functions.

Built-in Tools

Tool	Description	Example
get_crypto_price(symbol)	Fetches live crypto price from CoinGecko	"BTC" → "$67,420"
multiply_numbers(a, b)	Multiplies two numbers locally	144, 37 → 5328

Adding Tools

// In app.js, extend customTools object:
const customTools = {
  get_crypto_price: async (symbol) => { ... },
  multiply_numbers: (a, b) => a * b,
  // Add yours here:
  my_custom_tool: async (arg1, arg2) => { ... }
};

Model prompt includes tool schemas — it learns to call them automatically.

---

Project Structure

Project-Bob/
├── index.html          # Stark black/white UI
├── app.js              # Transformers.js runtime + tool engine
├── dataset.jsonl       # Fine-tuning examples (QLoRA format)
├── colab_training_notebook.py  # Google Colab training script
├── push_assets.py      # Deploy to GitHub Pages + HF Hub
└── README.md           # This file

---

Fine-Tuning Your Own Bob

Train a specialized version on your data:

1. Open Colab: Upload colab_training_notebook.ipynb
2. GPU Runtime: Runtime → Change runtime type → T4 GPU
3. Secrets: Add HF_TOKEN (write access)
4. Run All: ~15 minutes on T4
5. Result: Trained LoRA + ONNX 4-bit pushed to yummyfiles/Bob

Dataset Format (dataset.jsonl)

{"text": "<|begin_of_text|><|start_header_id|>user<|end_header_id|>What is the current price of Bitcoin?<|eot_id|><|start_header_id|>assistant<|end_header_id|><call_tool>get_crypto_price({\"symbol\": \"BTC\"})</call_tool><|eot_id|>"}
{"text": "<|begin_of_text|><|start_header_id|>user<|end_header_id|>Multiply 42 by 58 for me.<|eot_id|><|start_header_id|>assistant<|end_header_id|><call_tool>multiply_numbers({\"a\": 42, \"b\": 58})</call_tool><|eot_id|>"}

---

Deployment

GitHub Pages (Frontend)

python push_assets.py
# Pushes index.html, app.js → yummyfiles/Project-Bob (main branch)
# Enable Pages in repo settings: Source = main branch

Hugging Face Hub (Model)

python push_assets.py
# Uploads model artifacts → yummyfiles/Bob

---

Tech Stack

Layer	Technology
ML Runtime	Transformers.js 2.17+ (ONNX Runtime Web + WASM)
Model	Gorilla OpenFunctions v2 (3B, INT4)
Tools	Native JS async functions
Frontend	Vanilla HTML/CSS/JS (ES Modules)
Hosting	GitHub Pages (static)
Model Hub	Hugging Face Hub

---

Browser Support

Browser	Support
Chrome 94+	✅ Full
Firefox 93+	✅ Full
Safari 15.4+	✅ Full
Edge 94+	✅ Full

Requires: SharedArrayBuffer, WebAssembly, WebGL/WebGPU

---

Privacy

• No telemetry
• No external requests except:
  • Model download (once, from HF CDN)
  • get_crypto_price → CoinGecko public API
• All inference local — your prompts never leave your device

---

License

MIT — Free for any use.

---

Credits

• Transformers.js by Xenova
• Gorilla OpenFunctions by Gorilla LLM team
• Optimum for ONNX export
• CoinGecko for crypto prices

---

Built for local-first AI. No cloud required.