I like building things from scratch to understand how they work. Mostly neural networks and ML stuff, lately a chess engine in C++ too — but I'll automate anything that annoys me enough.
IxEngine: A chess engine in C++, from scratch — no libraries, just bitboards, alpha-beta search (PVS, null-move pruning, LMR, transposition tables) and a hand-tuned evaluation. Holds roughly even with Stockfish capped around ~2800 in blitz. My first proper systems project after a lot of Python, and probably the one I've had the most fun with.
Pion: Dense transformer LM. No gimmicks, just best practices — GQA, QK-Norm, SwiGLU, RoPE, tied embeddings, chunked cross-entropy. Pretrained Pion-small (103M) on RunPod, then fine-tuned on OpenHermes 2.5. Checkpoints on HuggingFace. (Still kinda bad, call it a W.I.P)
Tachyon 2: MoE transformer I built for pretraining. 8 experts per layer, only 2 active per token. Gets ~95k tok/s on an H100. The whole point was making something fast. (Slower than expected but functional)
Tokenizers: Tokenizers I use for my models.
Pretrained weights and checkpoints are on HuggingFace. Download them to try generation or continue training.
