Computer Vision · Security AI

How Real-Time AI Gun Detection Improved School Security

Angel Protection needed to protect hundreds of school cameras with firearm detection that actually worked — without crippling infrastructure costs or licensing constraints. We built it from the ground up.

Client
Angel Protection
Domain
School Security · CV
Start your project
AI gun detection with bounding boxes on a simulated school security camera feed
Angel Protection
Detection Accuracy
97.3%
near-zero false positives
97.3%
Detection accuracy
near-zero false positives
90%
Hardware cost reduction
same accuracy, fraction of cost
10×
More feeds processed
at 1/10th the previous cost
65%
Faster response time
via motion-activated scanning
The Challenge

Protecting schools required detection technology that simply didn't exist

Angel Protection operates AI-powered security systems across school districts. Their mission demands accuracy — a false negative isn't a missed sale, it's a missed threat. Existing solutions fell short on every front.

Inaccurate off-the-shelf systems
Existing solutions produced too many false alarms and missed real threats — unacceptable for school security.
No viable training dataset
No public firearm dataset suited the domain. Every training image required manual curation and validation.
AGPL-3.0 licensing blocker
Ultralytics (YOLOv8) uses a restrictive open-source licence that prevented commercial deployment.
Prohibitive hardware costs
Processing 100+ RTSP camera feeds with low latency demanded significant GPU infrastructure — at crippling cost.
The Solution

Built from scratch — licence-free, edge-optimised, massively scalable

Rather than patching an off-the-shelf system, Gradient Insight designed a bespoke distributed architecture: lightweight motion detection at the edge triggers centralised GPU inference — making it possible to monitor 100+ 4K streams simultaneously at near-real-time latency.

Custom, fully licence-free YOLO-based detection model trained from scratch on curated domain data

Distributed edge architecture: motion detection runs on-device; GPU inference centralised for scale

Domain-specific synthetic augmentation — lighting, angle, and occlusion variations baked into training

TensorRT-optimised inference pipeline handling hundreds of 4K streams per GPU node

Commercial Python package (angelcv) for camera-agnostic RTSP discovery and provisioning

~1-second end-to-end detection latency across 100+ simultaneous feeds

The Results

Outcomes that matter — at a fraction of the cost

The deployed system didn't just meet the original requirements — it redefined what Angel Protection's platform could offer. Accuracy, scale, and cost targets were all exceeded.

97.3%
Firearm detection accuracy
Validated across real-world school environments with diverse camera angles, lighting conditions, and partial occlusions — with near-zero false positives.
90%
Reduction in hardware costs
TensorRT optimisation and the distributed edge/GPU architecture eliminated the need for local GPU infrastructure at every site.
10×
More camera feeds processed
The same GPU budget that previously handled a handful of streams now powers over 100 simultaneous 4K RTSP feeds.
65%
Faster threat response
Motion-activated scanning and automated unwanted-person alerts cut the time from event to operator notification by nearly two thirds.
"Gradient Insight's solution transformed our approach. We now process 10x more feeds at 1/10th the cost while detecting threats traditional systems miss."
Lewis Matthews
Lewis Matthews
CEO · Angel Protection
Technology Stack
Custom YOLO ArchitectureTensorRTPyTorchEdge ComputingRTSP StreamingPython (angelcv)Distributed InferenceSynthetic Augmentation
At a Glance
Client Angel Protection
Industry School Security
Delivery Custom CV System
Latency ~1 second end-to-end
Feeds 100+ RTSP streams

Similar challenge?

We build bespoke AI systems for complex, high-stakes domains — from day one.

Book a free discovery call
"The architecture developed delivered outstanding threat detection accuracy while significantly reducing costs. Any company looking to push the boundaries of what's possible with AI would be fortunate to work with them."
Lewis Matthews
Lewis Matthews
CEO, Angel Protection
Ready to build?

Your domain, your data — built to your exact requirements

We don't adapt generic models. We architect bespoke AI systems that meet your performance, licensing, and infrastructure constraints — from scratch.