IndustryAuto Component Manufacturing
DisciplinesEdge AI · IoT · Computer Vision · Apps
Duration9 months · 6-plant rollout
RegionIndia (Pune, Chennai, Pithampur)

The problem

The supplier produced precision-cast brake components for global OEMs. Statistical Process Control sampled 1 in 50 units. Micro-fractures invisible to the naked eye were slipping through, surfacing as warranty claims 6-18 months later. Field-return cost in 2024 was a 9-figure rupee number, and one OEM had put them on probation.

The solution

A vision-AI inspection station at every line, running entirely on edge hardware (no cloud round-trip). NVIDIA Jetson Orin units pull 110 fps from 4K industrial cameras, run a custom YOLOv8-derived defect model, and reject parts in under 80ms. Defect images sync to a central data lake for model retraining; the line-side display gives operators instant visual feedback.

What we built

Edge inference at 110 fps

NVIDIA Jetson Orin Nano per line. INT8-quantised YOLOv8 derivative. 80ms p99 inspection latency, fully on-device.

Catches what humans miss

Detects sub-50µm hairline fractures, surface porosity, and dimensional drift — defects invisible to a trained QC operator on a moving line.

No cloud dependency on the hot path

Lines keep running even with internet down. Inspection results sync to the lake when connectivity returns. Critical for plants with patchy uplinks.

Operator HMI with live SPC

Touch-screen at every line shows the last 100 parts, defect heatmaps, drift trends, and one-tap reject overrides for genuine false-positives.

Closed-loop model retraining

Every false-positive operator override is collected, reviewed weekly, and folded into the next model release. Active learning loop closes within 14 days.

Per-OEM compliance pack

Per-shift, per-batch traceability with image evidence for IATF 16949 audits. Reduced audit prep from 3 days to 4 hours per plant.

How it’s built

Edge HardwareNVIDIA Jetson Orin Nano · 4K industrial cameras (Basler ace 2)
Vision ModelYOLOv8 derivative · custom backbone · INT8 TensorRT
Edge RuntimeNVIDIA DeepStream · ROS 2 · gRPC for HMI
Data LakeAWS S3 · Iceberg tables · Athena for SPC queries
TrainingPyTorch · Lightning · 1.4M annotated frames over 9 months
HMITouch-screen panels · Flutter desktop
MES IntegrationOPC-UA bridge to existing Siemens / Wonderware MES
ComplianceIATF 16949 · ISO 9001 audit-ready evidence

The numbers

99.4%
Defect detection recall vs. 2% under SPC
-78%
Field-return cost in year one
110 fps
Inspection throughput per line
80ms
P99 detection latency
6
Plants live in 9 months
14 days
Active-learning loop closure

“We went from being on OEM probation to winning a new platform from the same customer. The model literally sees what our most experienced inspector misses.”

— Plant Operations Director, Tier-1 Auto Supplier

Have a project that looks like this?

If your engagement combines 3 or more disciplines, we’d like to hear about it. Tell us the constraint, the deadline, and the outcome that matters — we’ll come back with a scoped proposal.