Inspiration
Every household has that moment: the dishwasher leaks, and now you are hunting through years of NAS folders for the model number, the receipt, and the warranty terms. The evidence exists — a phone photo of the label, a scanned receipt, a warranty PDF — but it is scattered across folders and impossible to trust quickly. We built Appliance Evidence Passport so a UGREEN AI NAS can answer one appliance question with evidence you can verify, not a guess.
What it does
Say "my dishwasher is broken" and name one authorized folder. The Skill runs a read-only scan, reads the model and serial from label photos (OCR and vision), extracts purchase facts from receipts (structured extraction), answers warranty questions from manuals and warranty documents with page citations, and groups duplicate evidence by checksum. It then shows a draft passport with VERIFIED, CONFLICT, and MISSING tables — every fact linked to a real source path and SHA-256 hash. After your approval it saves the passport Markdown and can copy the selected evidence into a per-appliance folder. Nothing is ever moved or deleted, and conflicting sources are preserved, never silently resolved.
How we built it
A Markdown-only Skill for the Uliya standard (no scripts, no external services). The parent orchestrates: read-only discovery first, a draft-and-approval gate, then two approved writes. Each step delegates to a registered NAS expert — file-agent (ug-filesystem, ug-filemgr), photo-agent (ug-photo), document-agent (ug-document), knowledge-agent (ug-knowledge) — passing only real upstream results downstream, handling all five delegation states (completed, failed, cancelled, waiting_user, waiting_approval), and treating empty results honestly. The package passes the official skill checker with zero errors and zero warnings.
Challenges we ran into
Designing for honest uncertainty: OCR misreads and conflicting documents are the norm, so the core design decision was to make conflicts and missing evidence first-class output instead of forcing a single answer. We also had to engineer every instruction around strict platform rules — no direct CLI access, approvals for every write, exact delegate formats.
What we learned
Evidence workflows need conflict handling, not just extraction. A passport that says "I don't know yet, here is what is missing" earns more trust than one that guesses.
What's next
Validate OCR and extraction quality on a real iDX6011 Pro, extend the passport to service history and repair costs, and let households track warranty expiry as a timeline.
Built With
- computer-vision
- linux
- markdown
- nas
- natural-language-processing
- ocr
- python
- ubuntu