01
AI detection & inference
Multi-model weighted-box-fusion ensemble — open-vocabulary, grounding, thermal, C-UAS, pose and behaviour — with motion gating and a global tile budget.
The core loop
Detect, alert, decide, dispatch, operate — the entire loop, on hardware you own.
↻ Closed loop · operator retains acknowledge, launch and abort at every stage
One frame, one decision
The loop, drawn.
Multi-model detection over a single feed, the alert grading it raises, the policy that decides whether a response is warranted, and the dispatch that gets written down. Four beats, in order.
DetectAlertPolicyDispatch
11 objects · one feedPolicy
Level 2 in a monitored sector, cooldown clear — a response is warranted
Dispatch
Nearest available drone routed to the camera’s GPS position · operator retains acknowledge, launch and abort
An illustration of the ICCS loop — detect, classify, decide, dispatch — drawn over a stock aerial photograph. Not a live feed, not a customer deployment, and not a screenshot of the product.
Capability inventory · 11 domains
Every domain shipped, not roadmapped.
Detection, sensors, drones, mapping, recording and access control are one system with one audit log — because in an accreditation review, five vendors is five arguments.
One platform, not five integrations
Detection, sensors, drones, mapping, recording and access control are one system with one audit log — because in an accreditation review, five vendors is five arguments.
Engineering detail
Sized, measured and deployable on hardware you already own.
Numbers from the running system, not a benchmark sheet.
At a glance
Scale and response
80 configured cameras on a single GPU node, with roughly 15–20 concurrently active 1440p streams — CPU-bound, not GPU-bound. Up to four autonomous drones over MAVLink and ArduPilot.
- Store · PostgreSQL 16 with SQLite as native development fallback
- Stack · Python / FastAPI · React / Vite · Ultralytics YOLO
AI model stack
Eight models, fused
Weighted Box Fusion across a primary detector, a transformer detector, open-vocabulary and grounding models, thermal and sliced-tile inference, plus a counter-UAS tracker. Approximately 28 GB VRAM for the full stack.
Security & deployment
Built for the accreditation review
Three-role RBAC, JWT authentication, a tamper-resistant audit log, dormant TOTP MFA and encrypted camera credentials — the controls an assessor asks for, present before they ask.
- Deploy · Native · Docker (CPU and GPU) · Windows via WSL2 · offline and air-gapped · systemd service
Weighted Box Fusion across eight models
≈28 GB VRAM for the full stackPrimaryTransformerOpen-vocabGroundingThermal / IRSliced-tileC-UASPose
Motion gating and a global tile budget keep the ensemble inside the compute envelope — which is why 80 configured cameras run on a single node without the GPU becoming the bottleneck.
Measured, in the field
What the detection stack actually returns.
Figures from the classified perimeter programme this engine was proven on. The client cannot be named; the numbers can.
97.8%
Detection accuracy
ARGUS, air-gapped
0.18%
False positive rate
Below the alert-fatigue floor
100%
Offline capable
Tested with the NIC down
94.3%
Acoustic accuracy at −5 dB
SENTINEL, sub-50 ms
Deployment targets