1 min read
Ames ensemble wins and failures
Fold safe preprocessing matters more than the fifth tree model.
Writing
Notes from the build.
Six short posts about the parts that break. Each one starts from a real failure mode in the work, not from a definition. Missing numbers stay marked as pending.
1 min read
BigQuery and Power BI: when the region is the bug
The query is fine. The data location and the dashboard region are not.
2 min read
ECG digitization: the grid is not the signal
A clean plot can still hide a bad calibration. Make the scale explicit.
2 min read
Elo for sports prediction: a baseline, not a finish line
Elo is cheap, strong, and easy to leak against. Treat it as the baseline.
1 min read
RAG lessons: evaluate retrieval first
A fluent answer is not evidence. The retriever is the part you can measure.
2 min read
Voice agent latency: the 800 ms budget
A voice turn is a latency budget. Measure each stage before tuning the prompt.