Field notes on machine learning, systems, and the quiet craft of building things that think.
Gradient descent is the most honest teacher I know. It will not tell you where the answer is — only which direction is less wrong than where you stand. For years I mistook that for a limitation. Now I think it might be the whole lesson: you do not need to see the valley to reach it, you only need the discipline to keep taking the next small, humbling step.
Read the essayA thousand Kaggle submissions taught me one thing the textbooks underplay: your cross-validation scheme is a model too — and an unexamined one will overfit your judgment long before your network overfits the data.
No single tree is right. That is precisely why the forest is. On bagging, boosting, and why the most reliable systems I have shipped were never built from a single confident voice.
Most outcomes have not happened yet — and pretending otherwise is how good models quietly lie. A field note on censoring, hazard, and modelling time-to-event without losing your mind.
The model is five percent of the problem. The other ninety-five is everything that has to be true for the model to keep mattering on a Tuesday at 3am. A pragmatist's case for boring infrastructure.
Predicting tomorrow's grid taught me to respect the gap between a good backtest and a real Tuesday. Notes from a season spent forecasting demand that refused to read the textbook.
A raw column is a fact; a feature is a sentence about what that fact means. The best feature I ever built was not clever — it was simply the question a domain expert had been asking all along.
New notes land between experiments — roughly when an idea has earned the patience to be written down.
Start a conversation