The team, the lesson, the working principles.
ORIGIN — LAGOS · 2024
The cameras already saw it.
A bank branch in Lagos had forty-eight cameras. After a robbery in early 2024, the footage was perfect. The men entered through the side gate at 02:14. They moved with practiced choreography. Every frame was on tape.
And every frame was useless. No one was watching at 02:14. The recording rolled until 06:00, when the morning guard checked the screens and called.
We started Sentinel because the next time it happens, the cameras should be the ones to call — before the getaway car leaves the curb.
Person-detection is not threat-detection.
The honest takeaway: detection without context is just noise dressed up. A person on a hotel staircase at 23:00 is a guest. A person on a jewelry counter at 23:00 is a different story. The same model, the same bounding box, two completely different calls.
So we stopped tuning the detector and started talking to the people whose job is to actually watch the screens — branch managers, hotel duty officers, jewelry shop owners, supermarket loss-prevention leads.
The result is the Behavioural Pattern Framework — 160 specific threats, organised by sector, each with its own evidence signals and its own response policy. The cameras are the eyes. The framework is the brain.
Six things we won't compromise on.
i.
Sector before signal
We don't tell you a person was detected. We tell you a guarded-vault perimeter was crossed at 02:14 and route it accordingly.
ii.
Edge first
Frames stay on your premises. Only metadata and incident snapshots leave the box. Your bandwidth and your privacy thank you.
iii.
Cryptographic identity
Every edge box generates a private key on first boot. The cloud verifies signatures, not tokens. Stolen credentials cannot replay.
iv.
Honest abstractions
If a model misclassified, the dashboard says so. If a camera went offline three minutes ago, the dashboard says so. No green checkmarks lying to you.
v.
Operator-led
Every template in our catalog was confirmed by someone whose job depends on catching it. We don't invent threats from a desk.
vi.
Built in Lagos, shaped for Africa
Intermittent power, low-bandwidth uplinks, hardware available locally — the platform was designed against these constraints, not retrofitted to them.
Two years. No shortcuts.
2024
First detection on a Lagos branch
Prototype YOLOv8 against one RTSP stream. Forty false positives in one night. The honest baseline.
2025 · Q1
Behavioural Pattern Framework
Sector taxonomy locked: 9 verticals, 16 sub-sectors, 160 codified threats. Built with operator input — every template signed off by someone on the floor.
2025 · Q2
Edge agent in production
PyInstaller binary, ONNX Runtime, RTSP capture, local SQLite. Signed-challenge handshake on every WebSocket reconnect.
2025 · Q3
Multimodal classifier
Claude vision wired into the alert pipeline. False-positive rate dropped 74% on field trials. The classifier writes the ruling.
2026 · Q2
Self-healing fleet
Auto-update with SHA-256 atomic swap, GitHub-Releases CDN. Every box updates itself; we sleep through the rollout.
ENTRY 04 — OPERATORS
Two people,
one tower PC, a lot of long nights.
We're Vergold — a small team out of Lagos building security products that make sense for the places we live in. Sentinel is our first commercial release. We're not a research lab and we don't intend to become one.
The team is intentionally small — a backend engineer, a frontend designer, two security-domain advisers from banking and retail, and a hardware partner who handles the field installs. We answer every support email ourselves.
— STAFF FILE
- Founding engineerBackend, edge agent, ML
- Founding designerFrontend, dashboard, brand
- Security adviser, Banking20 years on the operations side
- Security adviser, RetailLoss-prevention, multi-site ops
- Hardware partnerField deployment, edge box assembly
ADDENDUM
We'd rather be the right fit than the only choice.
If you run a site in one of the sectors we've modelled, we'll run a free assessment and tell you honestly whether Sentinel solves your problem.