Inspiration
Small NGOs and community projects must prove to donors how every dollar was spent. The proof exists, but it is scattered: contracts in one folder, bank slips in another, delivery photos on a phone. Auditors find the gaps months later: a payment with no contract, a delivery with no proof. I wanted a Skill that finds those gaps before the auditor does, without uploading sensitive finances or beneficiary photos to a cloud service.
What it does
Proof Trail verifies spending claims against evidence already stored on a UGREEN AI NAS. For one project folder and review period it reads contracts, invoices, payment slips, delivery photos and meeting recordings, and matches each expense to five proof slots: authorisation, billing, payment, delivery and corroboration. It reports what is proven, what is missing and what contradicts, for example an invoice for 200 units against a photographed delivery note showing 180, or one photo cited for two different expenses. After the user approves, it saves a binder with a prioritised fix list and a SHA-256 integrity manifest.
How I built it
Proof Trail is a pure orchestration Skill written to the Uliya Skill Development Standard: one SKILL.md and four linked reference files, with nine delegations to registered NAS experts (file, document, photo, knowledge, and optionally security-monitor). There are no scripts and no external APIs. Source files are strictly read-only, and the only write is the final binder, behind approval. A verify-before-report step re-checks every suspected contradiction with a second read-only delegation, so only confirmed inconsistencies are shown. I also built a synthetic rehearsal dataset with planted gaps and an answer key for testing.
Challenges
- Crowded ideas. My first idea, AI home inventory from photos, turned out to be crowded by many apps, so I moved to verification, which is less served.
- Delegation only. The Skill cannot call NAS functions directly, so every rule had to work as an instruction that an expert can follow.
- No NAS yet. I could validate structure with the official checker but not runtime behaviour. That is what the Finalist testing day is for.
What I learned
Search and question answering are already commodity features on NAS devices. The valuable part is trustworthy output: never inventing evidence, saying "not found" when proof is missing, and checking a finding twice before reporting it.
AI tools used
Claude (Anthropic) was used to research, design and draft the Skill, proposal and rehearsal dataset. The UGREEN Skill Developer Package checker was used for static validation.
What's next
Test on the assigned iDX6011 Pro, measure real run times, tune the instructions, and add playbooks for other audit frameworks.
Built With
- agent-orchestration
- ai-agents
- audit
- claude
- document-extraction
- evidence-verification
- exif-metadata
- grant-compliance
- local-ai
- markdown
- nas
- ngo
- ocr
- privacy
- sha-256
- skill
- ugos
- ugreen-ai-nas
- uliya
