Samit Mohan
Bangalore · ML engineer at Beltech AI since Jun 2025
≤10ms CV inference at 0.9+ F1 on city CCTV. I write and rebuild things from scratch.
Posts
all posts →- Mar 11, 2026building pytorch from scratch
- May 1, 2026inference engineering
- Mar 26, 2026what happens when you press 'submit' on chatgpt
- Mar 18, 2026from residual connections to attention residuals
- Apr 2, 2026the annotated microgpt
Projects
github →- mathLM — 40M GPT trained from scratch on math. Custom BPE, GQA + RoPE + SwiGLU, sparse MoE, SFT + GRPO reasoning RL. Evaluated on GSM8K.
- minitorch — PyTorch in ~1300 lines. Reverse-mode autograd, conv2d, optimizers, LR schedulers. 96%+ on 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.
- tennis-analysis — YOLO + ResNet on match footage. Court detection, ball tracking, shot classification. Tested on Wimbledon clips.