← All workCase 03 / 07
Kaggle · Data science
Competitive data science, run like research
33rd of 356 teams in a poker fraud-detection competition, where every submission had to earn its place on an honest holdout first.
- Role
- Research lead, with AI agents
- When
- 2026
- Status
- Top 10%
The problem
Competitions tempt you to tune toward the public leaderboard. It looks like progress and often is not: the public split rewards noise.
What I built
I built a local research toolkit and ran the competition like a lab. Every experiment is registered, and the measuring stick is an honest five-fold holdout, not the public leaderboard: ideas that only improved the public score were treated as noise.
AI agents run the experiments under these rules. Every submission needed my explicit approval, and every decision and finding is logged so nothing gets re-tried by accident.
- idea
- registered experimentlogged before it runs
- honest holdoutfive folds, not the leaderboard
- decision logwhat worked, what not to retry
- submit, with approval
Where it stands
33rd of 356 teams in Detect Suspicious Value Transfers in Poker (top 10%).
The approach has since carried into hyperspectral object detection on 16-band imagery, low-light object detection and a live memecoin pump-or-dump league, now in a workbench where every experiment writes down its expected result, and what would prove it wrong, before it runs.
- Python
- DuckDB
- LightGBM
- PyTorch
- Kaggle API