Edge AI + Human‑in‑the‑loop

Your team focuses on customers, ShrinkHalt AI + human reviewers handle loss prevention.

Skeleton‑based detection runs on affordable edge devices. Flagged events are verified by remote loss‑prevention specialists. Zero distraction for floor staff.

🛡️⚡
The reality

Retail shrinkage keeps growing, but most surveillance footage is never reviewed.

Store teams either ignore suspicious activity or waste hours watching cameras — time that should be spent on service and selling. Traditional AI alone generates too many false alarms.

📉

$112B+ annual shrinkage

Globally, retail shrinkage exceeds $100 billion annually. The majority of incidents happen in plain sight, yet real‑time capture remains elusive with legacy methods.

🕵️‍♂️

Less than 2% of theft events flagged

Standard CCTV review is reactive and understaffed. By the time footage is reviewed, the opportunity for intervention has passed.

⏱️

Associates lose hours on LP tasks

Manually tracking suspicious behaviour pulls staff away from the sales floor, hurting customer experience and revenue.

How ShrinkHalt works

Automation built for loss prevention, not more screen time.

Our lightweight AI filters 95%+ of routine floor activity. Only potential theft events reach our remote human reviewers. Your team receives verified, actionable alerts.

🧠

Skeleton‑based AI detection

Compact Temporal Convolutional Network analyzes pose sequences (no raw video) to classify actions into four operational classes: Floor activity · Consumption · Concealment in bag · On‑body concealment. Runs entirely on edge devices.

👥

Remote human validation

When the AI flags a potential incident, a trained remote loss‑prevention reviewer inspects the short clip. Only confirmed events generate an alert — eliminating false positives.

Real‑time verified alerts

Store staff receive discreet, verified notifications via mobile or store system. Because every alert is human‑confirmed, your team can respond with confidence.

🖥️

Edge AI + remote reviewer pipeline

Video processing stays local on low‑cost edge hardware. When the model detects suspicious motion, relevant clips are securely sent to our remote review team for verification — no continuous cloud upload.

96.6%
Weighted F1‑score (27.5k clips)
99%
Recall for consumption events
~95%
Fewer false alerts after human‑in‑the‑loop review
<0.3ms
TCN inference per frame (edge overhead)
Why ShrinkHalt

A different kind of loss prevention platform

🎯 Skeleton‑based, not pixel‑dependent

Robust to lighting changes, shadows, and background clutter. Tracks key body joints, extracting motion patterns rather than skin or objects.

🌀 View‑adaptive AI

Feature‑wise Linear Modulation conditions the model on real‑time viewpoint metadata, allowing adaptation to different camera placements without retraining.

✅ Hybrid AI + human accountability

The AI acts as a high‑precision attention filter; every critical decision involves a remote loss prevention expert. No unsafe escalations.

📦 Sub‑1M parameter edge model

Extremely lightweight TCN architecture runs on affordable edge devices. Pose extraction is the main workload — classifier adds negligible overhead.

Edge‑native architecture · 0.9661 weighted F1

The core AI is a Temporal Convolutional Network (sub‑1M parameters) with view‑adaptive conditioning, achieving 0.9661 weighted F1 on a challenging held‑out test set (27.5k real‑world pose sequences). Feature engineering uses skeleton‑based kinematics, enabling robust edge deployment on hardware like Jetson Orin Nano.

Four operational macro‑classes: Floor Activity, Consumption, Concealment in Bag, On‑Body Concealment. Optimized for loss prevention triage.

Process

How ShrinkHalt delivers verified alerts

1

Read & detect

Existing cameras feed into an edge device that extracts pose data. The AI continuously analyzes motion patterns and flags suspicious actions.

2

AI → remote review queue

Only potential concealment or consumption events are sent to our remote loss prevention team. Non‑critical floor activity is discarded immediately.

3

Human verification → store alert

Reviewer validates the event within seconds. Upon confirmation, a real‑time alert (with anonymised clip or metadata) reaches your store staff discreetly.

Works with your stack

Seamless fit for existing retail environments

Any ONVIF / RTSP camera Jetson Orin Nano / Xavier Intel NUC / x86 edge Store mobile alerting (SMS / app) Existing VMS integration API

Event‑driven clip transmission only when the AI detects suspicious activity — minimal bandwidth usage.

FAQ

Everything you need to know about ShrinkHalt

How accurate is the AI across different store layouts?
Our model was trained with geometric transformations and view‑aware kinematic features, achieving 96.6% weighted F1 on a scene‑held‑out test set. Performance generalizes across moderate viewpoint shifts; extreme occlusion may reduce precision.
How do human reviewers verify events?
When the AI flags a potential concealment or consumption action, a short clip (pose overlay or anonymised footage) is queued to our remote loss prevention team. Reviewers confirm or dismiss the alert within seconds. This hybrid design ensures near‑zero false positives for store staff.
Does ShrinkHalt require new cameras?
No. It works with standard IP security cameras. You only need a low‑cost edge device per store section (e.g., Jetson Orin Nano) to run the detection model. Existing infrastructure is fully compatible.
How quickly can we deploy?
Deployment is straightforward and typically fast thanks to our plug‑and‑play edge software and API‑first design. We'll work closely with your team to get you live with minimal friction — no lengthy custom training required.
What actions does the system detect?
Four operational macro‑classes: Floor Activity (normal browsing/picking), Consumption, Concealment in Bag (front/side bag), and On‑Body Concealment (pockets, hoodie, pants, outerwear). The taxonomy was optimized for loss prevention triage.
Is remote human review available 24/7?
Yes, our remote loss prevention team operates across multiple time zones. Average verification latency is under 15 seconds, with priority handling for high‑risk behaviours.