Inbox Triage Digest for UGREEN NAS

Inspiration

A NAS at home or in a small office quickly becomes a shoebox: scans, invoices, contracts, photos and exports pile up in shared folders with names like "NOWE" and "do_sort". Nobody has time to sort them, and the good search features only help if the files are already named and placed well. I wanted the NAS itself to do the boring first pass: look at every new file, understand what it is, and hand me one short, trustworthy note per day.

What it does

Inbox Triage Digest is a nightly triage pipeline for a messy NAS share. It discovers new files, removes exact and near duplicates, extracts embedded text, runs OCR on scans and photos, and classifies every document with a transparent, human-readable scheme (invoices, contracts, manuals, personal records, media, unknown). Each morning it produces a single digest note: what arrived, how it was classified, which folders are getting messy, and which archive moves it recommends. Nothing is ever moved or deleted without explicit user approval - the digest is a proposal, not an action.

How I built it

The whole pipeline is a UGREEN Uliya skill package. A SKILL.md orchestrates six built-in Uliya agents (file-agent, document-agent, knowledge-agent, photo-agent, media-agent and security-monitor-agent) through delegate commands, with a strict approval gate before any destructive step. Two reference documents define the classification scheme and the digest format, so the behavior is auditable and easy to tweak. The package ships with generation-time capability evidence for every step, so the Uliya runtime can verify at install time that each planned capability actually exists on the device.

Challenges I ran into

Making OCR robust for phone photos of paper documents, defining near-duplicate detection that does not merge two different invoices from the same supplier, and keeping the classification scheme small enough to be readable by a human while still covering real household and small-office traffic.

Accomplishments that I am proud of

The package passed the official Uliya skill checker with zero errors and zero warnings on the first validation run, and the entire flow is approval-gated: the skill never touches the archive without the user saying yes.

What I learned

That a good agent skill is mostly about evidence and guardrails, not clever code: explicit capability evidence, a strict classification scheme and an approval gate do more for trust than any model call.

What is next for Inbox Triage Digest

If selected for round 2, testing on real NAS hardware with real family and office shares, a digest scheduler with configurable quiet hours, and user-taught categories that grow the classification scheme from corrections in the digest.

Built With

• ai
• automation
• document-management
• nas
• ocr
• ugreen