Oops2SOTA

Inspiration

Most AutoML tools only tune parameters inside a fixed solution space. We wanted something closer to an ML research team: a system that investigates data, creates features, compares training strategies, learns from failed experiments, and decides what to try next.

That idea became Oops2SOTA—turning every “oops” into evidence for a better model.

What it does

Oops2SOTA is an autonomous ML system built as a seven-node LangGraph workflow:

Understand → Inspect → Engineer → Train → Evaluate → Revise → Submit

It performs safe data exploration, generates and tests feature transformations, compares model architectures and ranking objectives, and evaluates every candidate against the current best solution. When existing approaches are exhausted, its research layer introduces literature-backed directions.

How we built it

We combined LangGraph agents with deterministic ML components. DuckDB supports read-only data investigation, modular Skills provide reusable techniques, and PyTorch and NumPy power model training.

LLMs generate hypotheses and explain decisions, while code controls scoring, leakage protection, acceptance, and convergence. Each experiment is recorded to keep the process reproducible and auditable.

What we learned

The hardest part was preventing the agent from confusing noise with progress. We learned that autonomous ML requires more than a capable LLM—it needs strict contracts, safe tools, measurable experiments, and honest failure handling.