Inspiration
Most residential security setups force homeowners to stream private video footage to third-party cloud servers while paying recurring monthly subscription fees. With over 51 million households relying on home security cameras, privacy leaks and subscription fatigue have become massive everyday frustrations. I wanted to build a tool that uses local NAS hardware to handle security forensics completely offline, keeping family data secure at home without costing a dime in monthly fees.
What it does
Sentinel acts as an autonomous local security investigator on your UGREEN AI NAS. When an incident happens—like a package delivery dispute, an unknown visitor, or a perimeter breach—the app queries local camera feeds (ug-surveillance), resolves visitor identities, transcribes intercom or audio memos (ug-knowledge), and bundles everything into a clean, text-based markdown incident dossier saved directly to your local storage.
How we built it
The project was built using the official Uliya Skill Development Standard from the UGREEN developer package. It relies on a multi-agent orchestration architecture rather than a monolithic script. The parent Markdown instruction set (SKILL.md) safely delegates tasks to specialized registered sub-agents: security-monitor-agent for timeline and event extraction, photo-agent for snapshot verification, knowledge-agent for audio processing, and document-agent for compiling final reports.
Challenges we ran into
Designing the workflow without physical access to a UGREEN NAS during Round 1 required rigorous adherence to local state boundaries and error handling. Ensuring the system never hallucinates missing face IDs or unrecorded camera events when timelines come back empty was tricky, which required writing strict negative constraints and explicit empty-result checks into every step of the orchestration logic.
Accomplishments that we're proud of
Successfully structuring a complex multi-workflow security assistant that passes the strict local static checker (check-skill.mjs) with zero errors and zero warnings. Maintaining absolute data privacy by design—ensuring that biometric markers, video snapshots, and logs never leave the physical UGREEN enclosure.
What we learned
Building for an on-premise AI platform shifts the focus entirely from cloud reliance to local efficiency. We learned how powerful strict agent delegation and schema-backed tool mapping are for keeping local hardware processes predictable, secure, and user-approved.
What's next for Sentinel
Deploying and testing the package directly on a physical UGREEN AI NAS during Round 2 finalist access. We plan to expand automated object detection rules and refine the markdown layout of exported dossiers based on real-world camera feeds.
