Skip to content

Feature Matrix & Product Roadmap

Patrol Sight · HumanityAI'd — July 2026

Confidential

Download

PDF

1. Feature matrix (v1.1.7 — running today)

Capability Status Evidence / notes
Hands-free face recognition in the officer's lens ✅ Live ~68 ms typical glasses-to-label, measured stage by stage
Hardware H.264 streaming from AR glasses ✅ Live Camera2 → hardware encoder → RTP/UDP; 2560×1440 at 20 Mbps, 0% packet loss
SCRFD detection + ArcFace 512-d embeddings on TensorRT FP16 ✅ Live Model choice benchmarked across 0–90° yaw
GPU FAISS IVF-PQ watchlist search ✅ Live 93,728 persons in production; P95 8.5 ms over 1,931 searches
Two-layer tracking (spatial + identity cache) ✅ Live Search skipped on 97.6% of frames; GPU utilisation 85% → 40%
Colour-coded category labels in-lens ✅ Live Encoded name @@category@@ #colour format, 10 colours verified
Label stability (smoothing, stale eviction, duplicate suppression) ✅ Live EMA smoothing + spatial matching + IoU coasted-track suppression
Long-range mode (8–10 m) ✅ Live Raised detector resolution, reduced frame rate
Bulk watchlist enrolment ✅ Live 40–60 images/second
Live index rebuild + hot reload ✅ Live No restart, no retraining
Face-quality assessment ✅ Live Suppresses unusable frames
Control console — live device monitor ✅ Live Per-device fps, latency, face counts, annotated preview
Targets, labels, tracking history, device tracking ✅ Live C2 console
Alert publication on internal message bus ✅ Live For downstream command-and-control integration
Role-based operator accounts + audit trail ✅ Live Audit middleware + operator audit-log view
Privacy mode — send faceprint, not image ✅ Live 512 floats (~2 KB) vs image (~100 KB)
Per-device sessions, one-command provisioning ✅ Live Device-keyed sessions; deploy-glasses.sh
On-prem single-host deployment, 34-service stack ✅ Live One-command deploy/health/update/teardown
Boot recovery incl. GPU driver race handling ✅ Live systemd unit + Swarm state + GPU startup wrapper
Zero-trust internal messaging + encrypted console payloads ✅ Live Per-service bus authentication; AES-256-GCM
Backpressure-safe ingest ✅ Live Newest-frame-wins + frame-rate-capped worker
Browser-webcam test mode ✅ Live Validate recognition without glasses

2. Known limitations (honest)

  • Recognition envelope: reliable to ~5 m, maximum 6–7 m in good lighting; faces below 40×40 pixels are rejected rather than guessed. Long-range mode reaches 8–10 m at reduced frame rate.
  • One GPU server ≈ one site. The multi-GPU configuration is written into the stack file but not yet enabled and measured.
  • Concurrent-device capacity per GPU is not yet published as a measured number — it is bounded by decode and detection, not by search. We measure it during the pilot and contract on that figure.
  • The glasses transport is plaintext UDP, safe only on an isolated VLAN with no uplink. Encrypted transport is Phase 2 hardening.
  • Credential and network-surface hardening is incomplete — default credentials in the stack file, a committed API key in the app source, and datastore ports published on the host. Phase 0 of the hardening plan, and disclosed in full in the Technical Due-Diligence Report.
  • No third-party watchlist connector. The watchlist is the customer's own, enrolled through our console; national or interoperable database sync is roadmap v1.5.
  • Model licensing is unresolved for commercial deployment — the recognition weights derive from a research-licensed lineage. This is the top-priority pre-revenue item.
  • No alert back to the wearer other than the visual label — no audio or haptic channel yet.
  • No automated test suite or CI at stack and edge-app level.

3. Roadmap

v1.2 — Deployment-ready hardening (0–3 months)

  • Rotate all credentials; per-device provisioned tokens in device keystore; secrets out of source and out of the app package
  • Encrypted glasses transport (SRTP/DTLS or VLAN-level VPN); collapse non-nginx host ports to the private LAN IP
  • Fail loudly instead of silently falling back when the TensorRT engine cannot load
  • Remove the Docker-socket mount from the operations container; RBAC across console and engine APIs
  • Automated tests + CI; deployed-stack smoke tests; connection-pooler transaction mode
  • Resolve model licensing (commercial licence or replacement weights)

v1.5 — Operational reach (3–6 months)

  • Audio and haptic alerts to the wearer for high-confidence matches — eyes-up, ears-in
  • External watchlist connectors — sync identities from the customer's existing systems or authorised third-party lists
  • Command-and-control / VMS integration pack — publish alerts into the customer's existing operations platform rather than beside it
  • Multi-GPU enabled and measured; published concurrent-device ceiling per GPU class
  • Fleet management: device health, battery and thermal telemetry, over-the-air app update
  • Thermal-adaptive capture ladder for long shifts

v2.0 — Beyond identification (6–12 months)

  • Behaviour and anomaly analytics (loitering, crowd surge, restricted-area entry)
  • Multi-camera fusion: correlate the officer's view with fixed cameras for continuous tracking across a site
  • Vehicle and licence-plate recognition from the same wearable stream
  • Offline degraded mode: on-device matching against a cached high-priority subset when the link drops
  • Live operations dashboard with attention and coverage heat-maps

v2.x — Platform

  • Multi-site command console for ministries and estate groups
  • Federated watchlists with per-site scoping and cross-site audit
  • Evidential export packages formatted for judicial review
  • Accessibility and localisation of the officer and operator interfaces (Arabic-first console)

4. Differentiators to protect

  1. Measured sub-100 ms glasses-to-label — published stage by stage, not asserted. Nobody in this category publishes a latency budget.
  2. The two-layer tracker — eliding 97.6% of vector searches is what makes continuous patrol recognition affordable on one GPU. This is the core IP.
  3. On-premises by construction — recognition, watchlist and audit never leave the customer's network, which is where cloud recognition vendors cannot follow under GCC data-protection law.
  4. Customer-owned watchlist, no scraped data, human verification required — the governance posture is a competitive asset, not a compliance cost.
  5. A real operational envelope — one-command deploy, boot recovery, external volumes, release snapshots, zero-trust internal messaging. This is deployed infrastructure, not a demo.