When to use pre‑trained vs custom models?

Pre‑trained for generic tasks (OCR, basic tagging); custom for domain‑specific labels or high‑accuracy requirements.

How to measure production accuracy?

Ongoing sampling, updatable gold sets, assisted labeling and per‑class metrics (precision/recall/F1).

Dataset bias, privacy (faces/PII), temporal drift; mitigate with RAI policies, anonymization and monitoring.

Bring computer vision into business processes using Azure AI services.

Defect detection on production lines with object detection/segmentation.

Extract data from invoices, delivery notes and forms with validations and post‑processing.

Shelf counting, SKU recognition, stock estimation and anomaly detection.

PPE detection, restricted areas, safety zones and visual audits.

Heatmaps, paths, optimized layouts and promo monitoring.

Capture → Pre‑process → Model → Post‑process → Action/Alert → Feedback

Choose edge for latency/privacy; cloud for scale and centralization.

Option	Pros	Cons	When
Pre‑trained services	Fast go‑live, low upfront cost	Limited customization	OCR, basic tagging
Custom models	High accuracy on specific domains	Dataset & tuning required	Defects, proprietary SKUs
Hybrid edge + cloud	Latency/privacy + centralization	Higher complexity	Real‑time/sensitive requirements

Balance classes, vary conditions (lighting/angles) and ensure high‑quality labels.

Per‑class metrics, gold sets, A/B testing and frequent updates.

Privacy (faces/PII), audit, explainability and image‑rights management.

Which KPIs should we track?

Precision/Recall/F1 per class, false positives/negatives, latency P95, throughput and cost per frame.

How to handle drift?

Continuous sample collection, scheduled retraining, metric‑drop alerts and version control.

Do we need special hardware?

GPUs recommended for training; for inference evaluate GPU/CPU/edge accelerators based on latency and cost.