Computer-Vision Quality Control at an Indian Conglomerate — 96% Defect-Detection Accuracy on 14 Production Lines

MindMap deployed Quality Predictor (Qp) as the computer-vision defect-detection engine, Production Line Monitor (Pl) as the line-integration and operational-monitoring layer, ML Model Deployer (Ml) for the per-line model management, and Anomaly Detector (Ad) for the defect-pattern analysis.

Phase one was the defect-taxonomy and labelled-data build. The conglomerate's quality team worked with our delivery team to define the defect taxonomy for each production line — typically 12-24 distinct defect classes per line — and to build the labelled-defect dataset that the per-line models would train on. The labelled dataset combined historical-defect imagery from the legacy machine-vision and human-inspection records, with new high-resolution imagery captured specifically for the model training where the historical imagery was inadequate.

Phase two was the per-line model build. For each production line, a fine-tuned vision-transformer model was trained on the line-specific defect taxonomy. The base model was a Phi-3.5-vision variant chosen for its inference latency on the cost-optimised edge GPUs the line deployment used; the fine-tuning was per line, with the per-line model handling the line-specific product characteristics and defect patterns.

Phase three was the line-integration build. Each line's QC integration sits at the inspection-station with cameras capturing each unit at the required angles and resolution, with the model inference running on the per-line edge GPU and returning the defect classification within the per-unit cycle-time budget. The inspection outcome flows into the line's existing reject-and-routing mechanism (defective units routed to the rework station or scrap; passed units continue downstream).

Phase four was the operational-feedback layer. Quality engineers can review the per-unit inspection outcomes with the model-confidence and the imagery, with model-correction-feedback flowing into the per-line model continuous-learning loop. The platform's per-line defect-pattern analytics surface systemic defect causes for the manufacturing-engineering team's root-cause work.

The platform runs as an edge-and-cloud hybrid: per-line edge components handle the real-time inspection-and-classification with sub-cycle-time latency, while the central platform handles the model-management, the cross-line analytics, and the continuous-learning loop.

Quality Predictor's per-line model is a fine-tuned Phi-3.5-vision variant served on the line-side edge GPU (typically a single L40S per line). The model inference completes within 60-80ms per unit including the imagery pre-processing, fitting well inside the typical per-line cycle-time budget. The model is per-line and per-product-variant where the line produces multiple variants.

Production Line Monitor's line-integration layer handles the camera-and-sensor integration at each inspection station, the inspection-outcome flow into the line's reject-and-routing mechanism, and the inspection-rate-and-quality-rate monitoring for line-operations visibility. The integration uses each line's specific PLC and SCADA framework.

ML Model Deployer manages the per-line model fleet — model training (running in the central environment on the cumulative labelled-data plus the continuous-learning feedback), model validation (on held-out test sets per line), model promotion (canary deployment per line with rollback capability), and model monitoring (drift detection on the inference outcomes).

Anomaly Detector's defect-pattern analysis surfaces the systemic patterns — specific defect classes increasing in incidence (suggesting a root-cause shift in the line operations), specific defect-cluster patterns (suggesting equipment-condition issues), specific defect-temporal patterns (suggesting shift-and-operator variations) — for the manufacturing-engineering team's improvement work.