Inspiration
We are founding an affordable camera-and-drone security company in Nigeria, starting with large farms. In 2024, thieves harvested whole rice fields at midnight in Taraba State while outnumbered guards stepped back. Farms, estates, warehouses and schools already have cameras and guards — but nobody watches hours of footage, guard handovers are never checked, and only 25.7% of crimes in Nigeria are reported to police, often because there is no proof. We wanted the site's own NAS to tell owners what really happened.
What it does
Ask the NAS: "What happened at North Gate between 6 pm and 6 am?" Site Witness treats three sources as independent witnesses:
- Cameras – detected people, vehicles and unknown plates
- Guard reports – voice notes (transcribed) or handwritten logbook photos (OCR)
- Patrol photos – drone or phone shots checked by vision AI ("Is a fence broken? Is there smoke?")
Every finding is labelled Corroborated, Sources disagree – needs human review, or Unwitnessed (an offline camera is never reported as "all clear"), with camera snapshots attached. After approval, it saves the brief, translates it (e.g., into Hausa) and copies the evidence into an incident folder for police or insurers. It works for any time window, at any site.
How we built it
We built it with the official UGOS Skill Creation Tool. SKILL.md orchestrates five NAS experts — security-monitor, file, knowledge, photo and document agents — chaining real capabilities: camera event timelines and counts, stranger plates, audio transcription, vision OCR, vision Q&A, event evidence, save-text, translate and approved batch copy. Each step passes only real returned IDs, paths and text to the next. We used Claude to research Nigerian site-security realities and to run a multi-agent debate that stress-tested ideas against the judging criteria before writing a line.
Challenges we ran into
- Honesty over hype: Skills cannot send live alerts or fly drones, so we designed Site Witness as the investigator, not the alarm, and kept live alerting on our company roadmap.
- Fairness to guards: a contradiction could read as an accusation. Guards are treated as witnesses, never suspects, and disagreements are flagged only for human review.
- Privacy: unknown people stay "unrecognised", plates are only known or stranger, and nothing leaves the NAS — important under Nigeria's Data Protection Act 2023.
- Messy real data: different device clocks, accented speech and handwriting; we added time-alignment rules and an "unreadable" path instead of guessing.
Accomplishments that we're proud of
- A cross-examination approach to surveillance that surfaces disagreements and blind spots, not just event counts.
- A Skill that passes UGREEN's static checker with zero errors and zero warnings.
- Graceful degradation: it works with any two witnesses and says plainly which one is missing.
What we learned
The value of local AI is trust: evidence that stays on the site, works on poor rural internet (23% rural internet use), and admits what it doesn't know. We also learned UGOS's expert-agent model — delegate clearly, pass only verified results, and stop when data is missing.
What's next for Site Witness
- Test and tune on a real UGREEN AI NAS iDX6011 Pro.
- Pilot on Nigerian farms with our affordable solar cameras.
- Add our always-on layer beside UGOS Surveillance: instant SMS, WhatsApp and voice-call alerts, plus automatic drone-patrol photo intake.
- Expand from farms to estates, warehouses, construction sites, schools and city patrols, with one AI security brain per site that stays on site.
Built With
- ai-agents
- claude
- computer-vision
- edge-ai
- local-llm
- markdown
- multi-agent-orchestration
- nas
- ocr
- privacy
- security
- skill-md
- speech-to-text
- surveillance
- translation
- ugos
- ugos-skill-creation-tool
- ugreen-ai-nas
- uliya
- vision-language-model