MLOps platform for a 600-device edge AI fleet, case study.

Objectives

What the project needed to achieve

• Manage 600+ Jetson devices across 14 sites as one operated fleet
• Sign + verify model artifacts so only trusted models reach the field
• Cut median time-to-deploy a new model from 9 days to under 1 hour
• Detect accuracy drift automatically and roll back without human intervention
• Give plant operators a Foxglove-style dashboard rather than a black box

Challenge

Engineering constraint

The client had spent two years deploying 600 inspection cameras across 14 plants, but operating them as a fleet was a manual nightmare. Each model update required physically visiting plants. Accuracy drift on changing SKU mix went undetected until line operators escalated. A bad model that slipped through testing meant a 2-week recall across 14 sites. The team needed an operational platform that treated 600 edge devices as a managed service, not 600 individual deployments.

Approach

How Yantrix approached the work

1. 01Stood up an MLflow registry as the single source of truth for trained models, with strict promotion gates from staging → canary → production tied to evaluation-set metrics.
2. 02Built a signed-OTA rollout system — every model artifact is signed at promotion, devices verify the signature before loading, and the rollout pipeline targets canary devices first (5% of fleet) before expanding.
3. 03Implemented per-device drift detection — each camera logs feature-distribution statistics on a rolling window; the platform flags drift > 2σ from the canary baseline and either alerts or auto-rolls back depending on policy.
4. 04Built a fleet dashboard (FastAPI + React + Grafana) showing per-device latency, accuracy proxy, last deploy timestamp, drift status, and a one-click rollback action for plant managers.
5. 05Wrote runbooks and ran two operations bootcamps so the client’s plant IT teams could operate the platform without a Yantrix engineer on call.