All case studies
MLOps · Open Source

Sentinel: self-healing MLOps with physics-informed drift detection

A FastAPI microservice using physics-informed Z-score thresholds (>2.5σ) to detect data drift in production ML pipelines. Integrates an automated RAG agent powered by Gemini 2.5 for root cause analysis on error logs — no human escalation needed.

Drift threshold2.5σ
Response time<500ms
Automated RCA100%

The brief

Drift alarms that fire on every dashboard get muted within a week. The goal was a detector with a threshold anchored to each feature's own variance, and enough diagnosis attached that the on-call engineer knows where to look.

What we built

A FastAPI service that scores feature distributions against physics-informed Z-thresholds (2.5σ) and hands confirmed drifts to a Gemini 2.5 RAG agent that reads the error logs and writes a root-cause note beside the alert.

What it proved

Sub-500ms scoring, automated root-cause analysis on every triggered event, and no human escalations during the pilot window. The threshold is a per-feature knob set from its own history, not a global constant.

  • MLOps
  • FastAPI
  • Gemini
  • Docker
  • Anomaly Detection