Austin · Lisbon · KarachiTrust centredev@binarybreach.tech
BinaryBreach
Case study — Industrial manufacturer · multi-site

On-device vision inference on commodity hardware, with a central MLOps loop for drift detection and weekly model promotion.

SectorManufacturing & industrialDuration6 months to twelfth siteStackPyTorch · ONNX Runtime
Robotic arms on an automated production line
28msInference per frame
0.3%False-negative rate
12Sites in production
WeeklyAutomated promotion
The challenge

Where it started

Visual quality inspection was manual, inconsistent between shifts, and impossible to audit. A previous vendor's cloud-based system had failed because plant uplinks were unreliable and a round trip cost more time than the line allowed. Lighting and camera positioning differed at every site, so a single model trained centrally degraded badly on rollout.

The approach

  1. 01

    Deployed quantised detection models to commodity edge devices, running inference locally so a lost uplink degrades reporting but never stops the line.

  2. 02

    Built a central training loop that ingests sampled frames and operator corrections from every site to produce per-site fine-tunes from a shared base.

  3. 03

    Added drift monitors comparing live prediction distributions against the training baseline, alerting before accuracy visibly degrades.

  4. 04

    Automated weekly champion-challenger promotion with shadow evaluation and one-command rollback per site.

Stack
  • PyTorch
  • ONNX Runtime
  • NVIDIA Jetson
  • MLflow
  • Kubernetes
The outcome

Where it landed

Twelve plants run the system in production. False negatives — the expensive direction — sit at 0.3%, and per-site models are retrained and promoted weekly without an engineer visiting a factory floor.

Capabilities used

More work

A short conversation with an engineer, not a sales qualification call. If we're the wrong people for it, we'll say so and point you somewhere better.