kaggle-agent

An autonomous agent for Kaggle competitions: it profiles the data, freezes validation, runs experiments for hours, and submits only with my approval.

Role
Solo
Stack
  • Python
  • LightGBM
  • XGBoost
  • CatBoost
  • Optuna
  • PyTorch
  • SQLite
  • Streamlit
Links

Code for kaggle-agent

Problem

Most of the work in a tabular competition is careful bookkeeping: a validation scheme you don’t quietly break, experiments you can compare, and submissions you don’t waste. I wanted an agent to do the grinding while keeping the decisions that can go wrong (rules, submissions) with me.

What I built

Onboarding pulls metadata from the Kaggle API, profiles the files, detects the task and submission format, and runs integrity checks: target leakage, duplicates, train/test overlap and adversarial validation for distribution shift.

The validation scheme is chosen from the data and frozen. Folds are hashed, scores are only compared within one validation version, and a 10% holdout is hidden from every listing and agent.

A scheduler daemon keeps one GPU job and several CPU jobs running. A bandit over experiment “arms” (baselines, Optuna tuning, feature recipes, ensembling, calibration, seed bagging) picks what to try next. All state is in SQLite, so the loop survives restarts.

Submissions are prepared and validated automatically but need approval from an interactive terminal.

Results

  • Tabular classification and regression are complete; image classification has a working GPU training adapter.