Skip to content

Practice

Lead practice

Safety & Computer Vision

Production-grade computer-vision platforms for safety, security, and operational intelligence — deployed on the client's infrastructure.

AI-generated wide shot of a security operations centre — a wall of monitors displaying camera feeds and operators reviewing them calmly.

How it works

Safety-vision edge inference pipeline IP cameras feed a MediaMTX stream router, then a Triton plus TensorRT edge GPU inference engine runs computer-vision modules — face recognition, ANPR, weapon detection and PPE checks. Detections flow to an alert router and out to the operator-facing Ops Console. Side branches drop down to MinIO hot storage, a WORM audit log, and integration adapters into client systems. PRIMARY FLOW · EDGE INFERENCE < 100 ms IP Cameras Cam 1 · 2 · 3 RTSP / ONVIF MediaMTX stream router Triton + TensorRT NVIDIA edge GPU inference CV Modules FRS · ANPR Weapon · PPE Alert Router severity · escalation Ops Console operator UI PERSISTENCE & INTEGRATION MinIO storage 30-day hot retention WORM audit log tamper-evident Integration adapters event broker Client systems SMS · HR · parent comms key node
Edge-first inference at <100 ms. Per-site autonomous operation; the central NOC is a window into the site, not the brain.

What this practice is.

Our lead practice. A computer-vision platform for K-12 schools, universities, hospitals, manufacturing plants, retail chains, and corporate campuses. Twenty-one production-grade modules across identity, threat detection, and operational intelligence. RTSP ingestion, edge GPU inference via Triton + TensorRT, per-site autonomy through WAN outages, WORM audit logging. Per-module accuracy SLAs in contract. DPDP Act 2023 aligned. Currently in commercial discussion with our first multi-site engagement.

What we build.

Multi-site safety platforms

One platform engineered once, deployed across every site at marginal cost. Per-site autonomy, central NOC for cross-site rollups, no re-engineering per location.

Identity & access at perimeter

Hardware-grade FRS terminals with depth+IR anti-spoofing, ANPR for vehicle gates, RFID + face hybrid for high-security zones, visitor flows with parental consent.

Threat detection with human-in-loop

Weapon detection, intrusion, loitering, fight detection, audio analytics, PPE compliance, fire/smoke early warning — every P1 event verified by an operator before escalation.

Operational intelligence

Footfall, heatmaps, queue management, multi-camera tracking, crowd density estimation, predictive analytics — all aggregated, never per-individual.

How we engineer in this practice.

01

Edge-first inference

Real-time inference under 100ms on the client's GPU appliance. Per-site operations continue through WAN outages.

02

Honest per-module accuracy SLAs

Bands we commit to in writing, measured monthly, with service credits when we miss. No 99% claims that disappear after contract.

03

DPDP-compliant by design

Every engagement starts with a DPIA. Children's biometrics, healthcare records, and employee surveillance get the regulatory respect they deserve.

04

Hardware advisory by default, sales on request

Default model: we specify the hardware, the client procures it directly. No mark-up, no working-capital lock-in, no FX risk. When a client prefers single-vendor procurement convenience, we will also handle hardware sales — commercial terms agreed transparently in Discovery.

Stack in this practice.

  • NVIDIA Triton + TensorRT for edge inference
  • RTX PRO 6000 / L40S / L4 GPU appliances sized per site
  • MediaMTX for stream routing, MinIO for hot retention
  • React Ops Console, Keycloak identity broker, Kafka + Postgres for events

See the firm-wide stack →

What we won't build in this practice.

Emotion or mood recognition from face video

Scientifically discredited, banned under EU AI Act in education for good reason.

Mass surveillance against open watchlists

Only court-authorised or policy-authorised matching, with audit log and appeal route.

Automated disciplinary action without human review

Every consequential outcome needs a human in the loop and an appeal process.

Run a safety vision scoping conversation.

Tell us what you've already tried, what you've ruled out, and what success looks like. We come back within one working day.

Request a Discovery Call See our approach