Computer Vision / Edge AI Engineer: Real-Time Shoplifting Detection for Pharmacies (Veesion.com-like system)

About the project
We are a Montreal-based startup building a real-time, in-aisle shoplifting detection system for pharmacies, similar to Veesion.com. The system connects to a store's existing CCTV cameras, detects suspicious gestures (product concealment in clothing, bags or strollers), and sends a short video clip alert to staff on a mobile app within seconds.

We're looking for an engineer, or a small team, to build the full pipeline: edge AI inference, an in-store hardware module, and a mobile alert app.

Technical scope

Video ingestion: live streams from existing NVRs and IP cameras (RTSP / ONVIF)
Object detection: people, plus custom classes (product in hand, backpack, handbag, tote bag, stroller, store bag, jacket)
Multi-object tracking: persistent IDs for each person across frames and cameras
Pose estimation: skeleton keypoints to follow hand movements (hand to pocket, hand to bag, under clothing)
Action / behavior recognition: classifying gesture sequences into theft categories
False positive reduction: verification layer to confirm alerts before they reach staff
Edge deployment: in-store hardware module running inference on up to 30 cameras per store, optimized for real-time performance, with remote updates and monitoring
Data pipeline: labeling, data augmentation, and a retraining loop
Mobile app: iOS and Android, real-time push notifications with video clips, feedback buttons
Backend: API for alerts, stores, users, and clip storage

Deliverables

Phase 1: Paid test (1–2 weeks)

Pull a live stream from an IP camera
Detect and track people with persistent IDs
Send a push notification with a 10-second clip to a phone when a defined event occurs

Phase 2: MVP (8–12 weeks)

End-to-end pipeline running on an in-store hardware module, handling up to 30 cameras per store
Detection of 3 priority gestures: concealment under clothes, personal bag, backpack
Zone configuration per camera (shelf, checkout, exit)
Rules layer: severity tiers (high, medium, low), per-store toggles for each gesture, no duplicate alerts
Alert clip generation (a few seconds before and after the event)
Mobile app (iOS + Android): alerts feed, clip playback, "Theft / Not theft" feedback, "Declare an undetected gesture" button
Camera health monitoring: alert when a camera goes offline, is obstructed, or when the NVR has issues
Admin dashboard: stores, cameras, gestures, users
Retraining pipeline: staff feedback becomes labeled data for new model versions
Documentation and full source code handover

Phase 3: Pharmacy-specific gestures
Deblistering (opening packaging or removing security tags), shelf sweeping (bulk pickup into a bag), stroller concealment, store bag concealment.

Requirements

Proven experience deploying computer vision on live CCTV camera streams in production (please share a link or video)
Experience running AI models on edge hardware at scale (many camera streams on one device)
Experience with action recognition or pose-based behavior analysis
Understanding of false positive reduction in real-world environments
Privacy-by-design: no facial recognition, compliance with Quebec Law 25 / GDPR principles
Good written English (French is a plus)
Overlap of at least 3–4 hours with Eastern Time (Montreal)

Engagement

Contract, milestone-based payments. Possibility of a long-term role and equity for the right person.
All code, models and data are owned by [Company name] (work-for-hire).
Budget: [your range or "please propose"]

To apply, answer these 3 questions (applications without answers will not be reviewed):

Share a project where you processed live camera streams. What hardware, and how many cameras?
How would you reduce false alarms when a customer puts their own phone in their pocket?
What hardware and architecture would you propose to run detection on 30 cameras in one store, and why?