Blog

Retail: semantic catalog search and shelf computer vision

Tuning Top-N: How Many Candidates Should Go Into the Re-Ranker?

Re-ranker top-N tuning: too small and the re-ranker can only reorder a weak set; too large and you pay latency and cost on noise. Tune against recall at N.

Choosing an Embedding Model for Product Search: Small, Large, or Multilingual

An embedding model for product search: small is the default for an English catalog at scale, large adds a few points at ~5x cost, multilingual for non-English.

Computer Vision for Fresh and Perishable Sections: Where It Gets Hard

Computer vision for fresh produce retail works at bay level, not item level - no consistent packaging, variable appearance, fuzzy stockouts, and fast turnover.

Detecting Misplaced Products and Planogram Deviation, Not Just Gaps

Planogram compliance computer vision flags misplaced products and section-level deviation - it needs a machine-readable planogram and reliable product ID.

Sizing and Staffing the Human Review Queue for Ambiguous Detections

The computer vision human review queue takes the ambiguous confidence band - size it by detections times band width, and keep each review to seconds.

MQTT Quality of Service: Getting a Detection Event Out of a Flaky Store

MQTT QoS for retail edge: use QoS 1 (at-least-once) plus downstream dedupe for stockout alerts - QoS 0 can drop them, QoS 2 is overkill for idempotent events.

Fixed Cameras vs Staff-Phone Capture for Shelf Monitoring

Shelf monitoring camera setup: fixed cameras give consistent input and a tighter model; staff-phone capture is near-free to roll out but adds accuracy variance.

Confidence Calibration for a Detection Model: Making the Score Mean Something

Detection model confidence calibration: a raw 0.9 isn't 90% accurate, and confidence routing sets thresholds on that number. Fix it with temperature scaling.

Model Quantization for Edge Inference: Trading a Little Accuracy for Speed

Model quantization for edge inference converts 32-bit weights to fp16 or int8 - fp16 is near-lossless, int8 needs calibration and a measured accuracy check.

ONNX Runtime and Model Portability Across Mixed Store Hardware

ONNX Runtime for edge deployment: export the model once, run it through whichever execution provider matches each store's hardware - no per-device build.