Samit Mohan

ML engineer at Beltech AI. ≤10ms CV inference at 0.9+ F1 on city CCTV. I write and rebuild things from scratch.

Posts

mathLM — 40.4M-param GPT with GQA, RoPE, sparse MoE, SwiGLU, KV cache. SFT on OpenWebMath.
minitorch — PyTorch in ~1300 lines. Reverse-mode autograd, conv2d, optimizers, LR schedulers. 95%+ MNIST.
mini-code — Built to figure out what's actually inside an agent loop. Turns out it's 250 lines: tool calls, shell exec, streaming.
auto-research — LLM agents that generate, train, and validate PyTorch code in sandboxed subprocesses. Top-k search.
tennis-analysis — YOLO + ResNet on match footage. Court detection, ball tracking, shot classification. Tested on Wimbledon footage.

—