InferDyssey

A learning platform for exploring cutting-edge AI inference on Apple Silicon. Built with Reflex, powered by MLX, designed for hands-on experimentation with the Apple Neural Engine (ANE).

Quickstart

git clone https://github.com/yourname/inferdyssey.git
cd inferdyssey
chmod +x setup.sh && ./setup.sh
uv run reflex run

What it does

• Learn — Interactive explainer for the Apple Silicon inference stack (EXO, Flash-MoE, MLX, ANE)
• Vector Code Search — Semantic search across external repos using embeddings
• Workspaces — Run autoresearch experiments on 10M-124M param transformers via MLX
• Specs — Hardware detection → repo capabilities → trainable models → hypothesis builder

Requirements

• macOS with Apple Silicon (M1/M2/M3/M4)
• Python 3.11+
• OpenRouter API key (optional, for AI features)

Architecture

inferdyssey/
├── app.py              # Reflex entry point
├── views/              # UI views
├── core/               # Training, benchmarks, autoresearch, repo management
├── registry/           # Known repo capabilities (YAML)
├── data/               # Training data (shakespeare.txt)
├── experiments/        # Benchmark results (SQLite)
└── external/           # Managed external repo clones

Tech Stack

• Frontend: Reflex (formerly Pynecone)
• ML Framework: MLX (Apple's ML framework for Silicon)
• Vector Search: Integrating embeddings for code search
• Storage: SQLite for experiment tracking