A mature edge-AI surveillance platform: multi-camera detection, AI threat scoring, a live situational map, instant alerts, and evidence-grade reports. The detection models are trained for farms and critical infrastructure - and the same engine reconciles internal leakage in hotels.
VIGIL pulls RTSP streams from the field and runs object detection on every frame - people, vehicles, livestock, equipment - combining a fast detector with a vision-language model for the harder judgement calls. It doesn't just record; it recognises.
A goat at the fence at noon is noise. A person at the fence at 2am is a threat. Each detection gets a threat level from what it is, where it is, and when, so operators triage the few that matter instead of drowning in a thousand motion clips.
| Time | Camera | Object | Threat |
|---|---|---|---|
| 23:41 | CAM-03 Perimeter | Person, after hours | Critical |
| 22:58 | CAM-07 Store | Vehicle, unscheduled | Elevated |
| 21:14 | CAM-01 Pen | Livestock count −2 | Review |
| 18:30 | CAM-04 Gate | Staff, scheduled | Normal |
Cameras, sensors, and live incidents sit on a real map, so an operator running several sites sees where the trouble is at a glance - not a wall of feeds they have to watch all at once. Red pins pull the eye to what needs a decision now.
A mission groups the cameras, area, and objective for a site or an operation - a farm's night watch, an infrastructure patrol, a hotel audit window. Detections, media, and reports all hang off the mission, so the record of what happened stays together.
When a detection crosses the threshold, VIGIL sends an SMS to the people who can act - because the person who needs to know about a 2am intrusion is not sitting at a dashboard. Low-value events stay in the log; only the ones that matter interrupt someone.
An incident becomes a PDF report with the frames, timestamps, detections, and threat assessment attached - something an owner takes to the police, an insurer, or a staff review. The record is not a memory; it's a document, with an audit trail behind it.
A farm has no server room. So a camera runs on solar and an LTE router, tunnels out through CGNAT over a secure mesh, and VIGIL pulls and analyses the stream from a home PC or a VPS. The intelligence lives where the compute is; the site only needs power and a signal.
[Camera] ──RTSP──▶ [Solar + LTE router] │ (mesh tunnel, pierces CGNAT) ▼ [VIGIL @ home PC / VPS] ──▶ detect ▸ score ▸ alert Postgres · Redis · object detector · vision model
The same detection engine, trained and tuned for three very different jobs. This is what "the model is trained for farms and infrastructure, and used in hotels" means in practice.
Remote farms with no compute on site, watched from solar-powered cameras over LTE.
Substations, yards, stores, and remote assets that can't have a guard on every corner.
The engine compares what cameras saw against what the POS and PMS recorded, and flags the gaps.
VIGIL is a FastAPI platform with an async detection pipeline, a vision-language layer for hard calls, a Mapbox situational view, and edge-first deployment - production engineering, not a notebook demo.
A farm, a yard, a substation, or a hotel losing money to internal leakage - VIGIL is a mature build already running in the field. This page is a walkthrough with sample data.