VehicleWatch

The problem

A fleet has plenty of sensor data and almost no labelled failures. You can't train a classifier on breakdowns you haven't recorded.

What I built

1. 01

   Ingests sensor readings every 2 seconds and scores each one in under 100 ms on a single container.

2. 02

   Isolation Forest and Local Outlier Factor run together, because they fail differently.

3. 03

   Gemini turns each anomaly into a plain-English diagnostic.

4. 04

   Trained models persist in Redis, so scoring stays warm across restarts.

Honest limits

• No labels means no precision or recall to quote yet, and the alert threshold is tuned by inspection.