Skip to content
samit
An agent loopA pulse circles a think, act and observe loop; four tool calls fire from the act step on every lap.toolsthinkactobserve

Projects

Training pipeline, tokenizer through reasoning RLA run advances through tokenizer, pretrain, SFT and GRPO stages while a loss curve draws downward beneath them.40M params · from scratchtokenizertokenizerpretrainpretrainSFTSFTGRPOGRPOlosssteps →

mathLM

PyTorch

A small GPT trained on maths from scratch, tokenizer through reasoning RL.

source ↗
Reverse-mode autodiffValues flow forward through x squared and three x into y, then gradients flow back along the same edges and sum at the fork.reverse-mode autodiffx = 2leafx² = 43x = 6y = 10seed 1.0x.grad = 74 + 3forward: valuesbackward: grads

minitorch

NumPy · CuPy

Reverse-mode autograd in a few hundred lines of NumPy, written to understand backward.

source ↗live ↗
An agent editing a fileA prompt arrives, the agent reads a file, edits it, and reports back, one line at a time.minicode · ~350 lines› add binary searchReadbinary_search.pyEdit+ bisect_left()Bashpytest — 4 passeddone. leftmost insertion point.

mini-code

Python

A coding agent in ~350 lines, to find out what is inside the loop.

source ↗
Tracking a rallyCourt keypoints lock on one by one, two player boxes follow their ends of the court, and the ball travels a rally arc with its interpolated path drawn behind it.P1P2118 km/hcourt keypoints · player + ball tracking

tennis-analysis

PyTorch · OpenCV

Tracking players, the ball and court keypoints through match footage.

source ↗