The fifth entry in a live builder's log. A parameter sweep handed me a new best score. Three red flags said it was noise wearing a knob's clothing, so I refused to ship it. Most of competitive ML is declining to fool yourself.
The fourth entry in a live builder's log. I built a faithful local scorer, then fed it a biased test set, and it confidently signed off on a change that made me worse on the real leaderboard. A dev set built from failure cases certifies rescues, not deployments.
The third entry in a live builder's log. My first Kaggle submission was rejected before it ever scored, thanks to a stale sample_submission file and a Google Drive download limit. The plumbing turns out to be half the competition.
The second entry in a live builder's log. The Hyperspectral Object Tracking Challenge scores you on an unlabeled test set, so the leaderboard can't tell you your own score. Here's how I built a local scorer byte-faithful to the official one, and why measurement is half the battle.
The first entry in a live builder's log. I'm competing in the Hyperspectral Object Tracking Challenge 2026: track one object through video shot in colors the human eye was never built to see. Here's what the problem is, why it's hard, and where I'm starting on the leaderboard.