Back to work

Autonomous experiment loop

Autoresearch

A research agent needs a tight loop: edit, train, measure, keep or discard, and leave a reviewable trail.

What I built

Local experiments around Karpathy's autoresearch setup: a small LLM training harness where agents modify the training file, run fixed-budget experiments, compare validation loss, and iterate.

Stack

PythonPyTorchuvLLM agents

Status

Fork/reference workflow; phrased carefully

The local checkout tracks karpathy/autoresearch. This portfolio describes experimentation around the workflow, not authorship of the upstream project.