CASE STUDY · COMPUTER VISION · EDGE AI

Real-time shelf intelligence,
at retail scale

Empty and misplaced shelf stock quietly costs retail billions in lost sales. Earlier in my career I engineered production computer-vision systems that watch shelves continuously, detect out-of-stock, low-stock and misplaced items in real time, and alert staff before a customer meets an empty slot.

LATENCY <200 ms THROUGHPUT 15–20 FPS MODELS YOLOv7/v9 · TensorRT MATCHING LightGlue · multi-camera SERVING hot-reload · CI/CD

This page is an original, interactive illustration built by me, running on synthetic data generated in your browser. It contains no employer or client software, data, imagery, branding, or product names. Quantified specs above are from my public resume.

LIVE SIMULATION · SYNTHETIC DATA AISLE 03 · 60 SLOTS · 2 VIRTUAL CAMERAS

100%ON-SHELF AVAILABILITY

— msMEAN DETECTION LATENCY

0OPEN ALERTS

0DETECTIONS THIS SESSION

--:--:-- system ready. press RUN.

UNDER THE HOOD

How a system like this works

01 · CAPTURE

Shelf-mounted cameras

Fixed cameras face the opposite shelf, each covering many slots. Sessions are sensor-driven, so compute is spent only when something changes.

02 · DETECT

Optimised detection

YOLO-family detectors compiled with TensorRT run at 15–20 FPS, classifying every slot: stocked, low, empty, or wrong product.

03 · RESOLVE

Multi-camera matching

LightGlue feature matching and calibrated geometry reconcile overlapping views, so one physical slot never becomes two conflicting answers.

04 · ACT

Serve & alert

Hot-reload model serving pushes events to staff devices in under 200 ms, and replenishment time is tracked from alert to restock.

WHAT I ACTUALLY BUILT

The engineering behind the demo

Multi-camera shelf-vision pipelines on YOLOv7/v9 with TensorRT acceleration and LightGlue feature matching — sub-200 ms responses at 15–20 FPS in production retail environments, with high-precision camera calibration.
Zero-downtime ML serving — hot-reload model APIs (Flask/Gunicorn) integrated with event streaming and document storage, fully containerised with observability built in.
End-to-end CI/CD for models, so detection improvements shipped like software, not like science projects.
The same discipline, upstream and down: the harder problem is rarely detection — it's calibration, definitions ("what counts as low?"), and making alerts trustworthy enough that staff act on them. That thread runs through everything I build, including my current enterprise AI work.

Deploying AI where it has to work?

I'm a forward-deployed applied AI engineer — GCP Professional Cloud Architect and Anthropic Claude Certified Architect.

Contact on LinkedIn More work

Real-time shelf intelligence,at retail scale