All Case Studies
ManufacturingIndustrial Robotics2025

Aurora Vision Platform

Built a real-time defect detection system processing 4K video at 60fps across 12 production lines, replacing manual QA with sub-second automated triage.

Client
Industrial Robotics
Sector
Manufacturing
Duration
22 weeks
Team
4 engineers · 2 researchers
The Challenge

The starting point.

The client's existing QA process required two operators per line and missed an estimated 18% of defects. Latency budget for any automated solution was 200ms end-to-end.

The Approach

What we built.

  1. 01
    Collected and labeled 240k defect images across 9 categories
  2. 02
    Distilled a YOLOv9 detector into a 14M-parameter student model
  3. 03
    Deployed via NVIDIA Triton on-prem with hardware-accelerated preprocessing
  4. 04
    Built a feedback loop letting line operators flag false positives in one tap
Results

What shipped, what changed.

99.4%
Defect recall
47ms
End-to-end latency
12
Production lines live
$2.8M
Yearly scrap reduction
Stack

What's running in production.

PyTorchYOLOv9TritonTensorRTKafkaGrafana
Next case study
Lumen Document Intelligence