projects

OpenAI-compatible vLLM layer that trains speculative draft models during idle GPU time, improving accepted tokens per verifier step by 13.4% on Qwen3-8B.

DINOv3 in dependency-free pure Python — hand-written reverse-mode autograd and ViT, EMA teacher-student distillation, DINO + iBOT objectives, KoLeo regularization, axial RoPE with register tokens, and Gram anchoring.

High-concurrency math competition platform for 150k+ users across 11 partner organizations, with React/Node.js auth, submission, and leaderboard workflows. $20k raised.

PyTorch reproduction of "Deep Mixture of Experts via Shallow Embedding," recovering the paper's accuracy-vs-compute tradeoff on CIFAR-10.

Paper-to-GitHub matching pipeline across 500k+ ML/AI papers, surfacing research that still lacks a public implementation.