Sovereign Watch
One-Line Pitch
An AI that watches your street the way a sharp neighbour does — not just recording, but understanding what it sees, remembering what's normal, and knowing when something is wrong.
The Problem
South Africa has the neighbourhood watch. It has armed response. It has WhatsApp security groups where someone posts "suspicious vehicle on Protea Street" at 11pm and everyone squints at a blurry photo in the dark.
The watch is willing. The technology has always been the problem.
CCTV records. It does not think. It stores terabytes of footage nobody watches until after the break-in. The camera sees everything and understands nothing.
What you need isn't more cameras. You need one camera that thinks.
What Freddy Sees
Freddy Eyes is a vision system that runs on top of a live camera feed. It doesn't capture — it analyses, narrates, and remembers.
Every few seconds, Freddy looks at the current frame and asks: what is happening here, and has anything changed?
It builds descriptions like a person would:
"A man in a dark hoodie has been standing at the gate on the left side for the past four minutes. He is not moving toward the entrance — he appears to be watching the street. No vehicle is present. This is different from the last 40 minutes, which showed no pedestrian activity."
That's not an alert algorithm. That's situational awareness.
When the scene is quiet, Freddy doesn't just sit idle. He paints what he sees — scene interpretations, data infographics, monochrome terminal art, botanical portraits — all generated live from the camera feed and displayed on any connected screen.
The Stack
| Layer | What It Does |
|---|---|
| Camera | Tapo C210 pan/tilt IP camera — positioned at street or gate level |
| Vision Loop | Captures JPEG frames every 5s, analyses with Gemini AI every 8–90s (adaptive) |
| Memory | Remembers the last 20 high-significance moments — knows what "normal" looks like |
| Vibe Scoring | Rates each frame 1–5 for activity level — idle street vs active scene vs notable event |
| PTZ Control | Pan/tilt from any device — point the camera at what you want to examine |
| Display Surface | Any Android TV or phone shows the live feed + AI narration in real time |
| Autonomous Scouting | When the scene is static too long, Freddy pans around on his own — exploring, checking angles, then returning home |
| Director System | Freddy runs the TV view like a director — auto-cutting between live camera, AI artwork, and memory card replays |
| Art Generation | Four styles: scene interpretation, data infographic, monochrome terminal art, botanical portrait — all from live camera |
| Scene Context | Named entity memory. "That plant is Moose" — Freddy remembers forever and uses the name in every future analysis |
| Voice PTZ | Chat commands like "look left" or "camera home" move the camera in real time from /studio chat |
The South African Angle
Neighbourhood watches here already work because people are motivated. What they lack is:
- Eyes that don't sleep — the 3am shift nobody wants to do
- Memory across incidents — the same car that drove past slowly last Tuesday and the Tuesday before
- Language that matches the threat — not "motion detected" but "vehicle has stopped outside number 14 for the second time tonight"
- A shared view — everyone in the WhatsApp group seeing the same live feed, same AI commentary, same moment
What's Live Right Now
This isn't a pitch deck. The core system is running today on real hardware, watching a real space:
Live PTZ Camera
Real-time pan/tilt/zoom with AI vision analysis running continuously
Adaptive Analysis
Watches harder when something changes (8s), rests when quiet (90s cruise mode)
Vibe Scoring
1–5 activity rating per frame with memory ring buffer of notable moments
Android Surface
Any TV or phone shows live feed + narration via Kronos Surface APK
Autonomous Scouting
Freddy pans the camera on his own when idle — exploring and checking angles
Director Mode
Auto-cuts between live camera, artwork, and memory replays — no human needed
Four Art Styles
Scene interpretation, infographic, monochrome terminal, botanical — all from live feed
Voice PTZ
"Look left" in chat moves the camera. "Camera home" returns to start position
Scene Context
Named entity memory — "That plant is Moose" persists across sessions forever
Live Camera feed active Live AI analysis running Live Director system Built Surface APK
Roadmap
- Alert Push — WhatsApp or SMS when vibe score hits 4–5: "something notable is happening on your street"
- Multi-Camera — more nodes, more streets, all feeding one brain. Each Pi node is sovereign — runs locally even if internet drops
- Community Dashboard — a shared screen for the street WhatsApp group with live AI commentary
- Incident Log — searchable history of what Freddy noticed and when, exportable for SAPS reports
- KronosID Integration — community members join via
join.kronosfleet.net, identity managed atkronosfleet.net - NFC Guard Patrol — DESFIRE EV3 cards for physical guard checkpoint verification. Tap a card, the system logs location + time cryptographically. Guards can't fake a patrol. Built on KendoCards→ infrastructure.
Who This Is For
| Audience | Why Sovereign Watch |
|---|---|
| Residential estates | Existing neighbourhood watch + the 3am shift AI handles for you |
| Street communities | WhatsApp group gets a shared live AI view — everyone sees the same moment |
| Body corporates | Add AI oversight to existing CCTV without ripping out hardware |
| Security companies | Augment human guards with AI that never sleeps + cryptographic patrol verification |
| Any community | Anywhere with cameras that wants eyes that understand, not just record |
Connections
- Stochastic Gate→ — the memory system that makes "what's normal" possible
- KendoCards→ — identity infrastructure for community members and NFC guard patrol
- Zero Darkness→ — the broader ambient intelligence layer this fits inside