Tesserack
AI learns to play Pokemon Red entirely in your browser. WebGPU LLM + TensorFlow.js neural network + WASM emulator. No servers required.
Tesserack
Browser-based AI that learns to play Pokemon Red. No server required.
Overview
Tesserack combines an LLM (Qwen2.5-1.5B via WebGPU), a trainable policy network (TensorFlow.js), and a GameBoy emulator (binjgb/WASM) to play Pokemon Red entirely client-side.
Objectives are sourced from Prima's Official Strategy Guide (1999) — 47 ordered checkpoints from Pallet Town to the Hall of Fame.
Requirements
- Chrome/Edge 113+ (WebGPU)
- Pokemon Red ROM (.gb)
Quick Start
git clone https://github.com/sidmohan0/tesserack.git
cd tesserack/svelte-app
npm install
npm run dev
Architecture
| Component | Technology | Purpose |
|---|---|---|
| LLM | WebLLM (Qwen2.5-1.5B) | Action planning |
| Policy Network | TensorFlow.js | Learned action selection |
| Emulator | binjgb (WASM) | Game execution |
| State | Direct RAM reading | Ground-truth game state |
| Curriculum | Prima Guide (1999) | Structured objectives |
Storage
All data persists locally: experiences and model weights in IndexedDB, game saves in localStorage, LLM weights in browser cache (~1.5GB).
Links
License
MIT
Built by Sid Mohan
