newBrain

The problem

Understanding modern language models deeply means building one end to end — tokenizer, architecture, training and fine-tuning — rather than only calling an API.

What I did

Implemented a causal GPT transformer and its BPE tokenizer from scratch in PyTorch, added MLX LoRA fine-tuning, and wrapped it in a FastAPI/Three.js workspace with RAG and multi-agent chat.

Highlights

• Custom BPE tokenizer and a causal GPT transformer implemented in PyTorch (RoPE, SwiGLU, SDPA).
• LoRA fine-tuning on Apple MLX.
• FastAPI + Three.js workspace with RAG (ChromaDB) and multi-agent chat.