Research
Technical writing from the Data Pigeon team — methodology, field findings, and model decisions explained plainly.
The best failure predictor we found wasn't an error code. It was chargers getting slightly sloppy about when they reported.
Predict "no fault" every time and you'll score 97–99% accurate and catch nothing. Accuracy is the wrong metric here.
We tried to catch every failure with one model. Here's why that failed and what we built instead.