Not benchmark demos. Deployed systems: a surveillance AI watching 8+ simultaneous camera feeds at 91% detection accuracy, and a wearable-sensor pipeline classifying human movement at 92%+. You can watch the live demo videos on this site.
Manufacturers and facility operators who want visual inspection or monitoring without a research team
What you get
Object detection & tracking
YOLOv8-class detectors with multi-object tracking, tuned to your footage, not stock weights. My CCTV system cut false positives by 35% with per-zone threshold tuning.
Real-time video analytics
Multi-camera ingestion (RTSP/FFmpeg), on-stream inference, and WebSocket alert delivery to operator dashboards, 8+ simultaneous feeds in production.
OCR & document intelligence
Text extraction from scans, photos, and industrial displays using Tesseract, OpenCV, and deep learning, a service line I’ve delivered for international clients since 2020.
Pose & motion analysis
Human movement understanding from video and sensors: gait analysis and kinematics at 92%+ movement-classification accuracy in a clinical rehab platform.
Edge & cloud deployment
ONNX-optimised inference (2-3× faster on CPU), Dockerised serving, and the edge-vs-cloud tradeoff analysis for your latency and bandwidth budget.
A security operation needed automatic detection of crowd surges, loitering, intrusion, and fights across many cameras, without an operator glued to every screen, and without alert fatigue.
Built
YOLOv8 detection + multi-object tracking with a custom anomaly engine (per-zone dwell, red-zone entry, crowd density), snapshot evidence per alert, and WebSocket push to an operator dashboard. Evolved over four phases from prototype to production MVP.
On well-scoped problems with decent footage: 85-95% out of the gate, improving with tuning on your data. My production surveillance system runs at 91%. Anyone promising 99% before seeing your cameras is selling, not engineering. Accuracy is measured on your footage, in your lighting, at your angles.
How much training data do we need?
Often less than you think. Modern detectors fine-tune well on hundreds (not millions) of labelled examples, and for common object classes we start from strong pretrained weights. The audit includes a data assessment before you commit to anything.
Edge or cloud?
Edge when bandwidth, privacy, or sub-100ms latency demands it; cloud when you need fleet-wide models and easy updates. I’ve shipped both: ONNX-quantised models on constrained hardware and multi-stream GPU inference in the cloud, and will model the cost of each for your camera count.
Can you work with our existing CCTV/IP cameras?
Almost certainly. If it speaks RTSP (nearly every IP camera does), the ingestion layer I use (FFmpeg-based decoding into a frame buffer) handles it without hardware changes.