NasonNation // Command

A full-stack, self-hosted computer vision, biometric identification, and perimeter security platform architected to eliminate the false alarms, cloud latency, and privacy leaks of commercial NVR systems.

System Architecture: Split Edge-CPU + LAN-GPU Pipeline

Running a 7-billion-parameter Vision-Language Model (VLM) 24/7 on every video frame wastes power, while lightweight edge models alone hallucinate on complex lighting and inanimate objects. Command solves this with a two-tier distributed architecture:

1. Tier 1 — Always-On Edge Node (bossman — Debian Linux / Docker):
   • Sub-Second Video Fabric (MediaMTX): Ingests multiple 2K RTSP camera streams over TCP and serves sub-500ms WebRTC (WHEP) streams to desktop and mobile viewports.
   • 2-Pass Kinematic Displacement Engine (YOLOv8s + MOG2): Combines background subtraction with a custom 2-pass spatial tracker and Parked Vehicle IoU Footprint Locking, ensuring stationary vehicles and swaying foliage never trigger false alarms.
   • Real-Time Biometric Identification (YuNet + SFace): Aligns 5-point facial landmarks and extracts normalized 128-dimensional biometric embeddings in <3ms, matching against PostgreSQL vector galleries via cosine similarity (>= 0.363).
2. Tier 2 — On-Demand LAN GPU Co-Processor (NVIDIA RTX 3060 Ti):
   • Whenever the Edge Node verifies a true moving target, it dispatches the high-resolution crop over the local gigabit LAN to an RTX 3060 Ti running Qwen2.5-VL:7b.
   • The VLM performs semantic verification—purging inanimate false positives, extracting vehicle paint/body types, reading License Plates (ALPR) and Delivery Carrier logos (AMAZON, UPS, FEDEX, USPS), and writing a tactical summary back to PostgreSQL.

---

Key Engineering Highlights

• Zero-PII Threat Intelligence & FBI/DOC Watchlist Daemon: A background async worker paginates through the FBI Wanted REST API (Portland & West Coast field offices) and Oregon DOC Most Wanted feeds, computing 128-D SFace vectors purely in RAM (0 KB disk image retention) for instant watchlist alerting.

• Auto-Clustering “Repeat Visitor” Dossiers: Unknown visitors with high-confidence facial landmarks (>= 75%) are automatically clustered into persistent visitor profiles (VISITOR_01, VISITOR_02) with sighting counters and timestamps.
• OpSec-First OSINT Launchpad: One-click reverse-face search integration (PimEyes, Yandex) with an automatic Whitelist OpSec Lock that disables external uploads for trusted family members and strips all EXIF metadata.
• 5-Second H.264 MP4 Instant Replays & 24h Thermal Heatmaps: Rolling pre/post-trigger frame ring buffers encode 5-second H.264 clips for person and animal detections, paired with a live 16:9 SVG thermal dwell-time heatmap overlay.

View Code