Inspiration

My NAS holds years of scanned bills, contracts and phone photo bursts. Cloud AI would mean uploading private papers, and a normal AI agent with file access can be steered by text hidden inside a document. I wanted the NAS's own AI to do the sorting, without ever letting a file talk it into anything.

What it does

Private AI Curator uses UGREEN AI NAS's own document and photo AI to file documents by what they say and keep the best shot of every burst, with a preview, exact approval and one-step undo.

  • curate: document-agent labels each document from a closed list of six labels; the Skill files it by label and year.
  • best-shot: photo-agent groups bursts and rates each frame; the keeper stays, the rest go to a Curator Review folder (never deleted).
  • dedupe and undo run: exact duplicates only, and every run is reversible from a ledger.

How we built it

A Markdown-only Skill built to the Uliya Skill Development Standard (SKILL.md plus five linked references), orchestrating file-agent, document-agent and photo-agent through file-workflow, ug-filesystem, ug-filemgr, ug-document and ug-photo. The AI experts only get narrow, read-only questions and must answer in strict JSON (a label with high confidence, or 0–10 scores). The Skill builds every destination path itself, a policy gate applies scope rules and caps that can only be tightened, and the preview ends with an exact approval line like APPROVE 1 MOVES 19 DELETES 0.

Challenges we ran into

There is no snapshot capability in the catalog, so undo is built on the recycle bin (del → final_path → restore) plus reverse moves, with sha256 checks before restoring and clear reasons for any conflict that is skipped.

Accomplishments that we're proud of

The model never chooses a path or an action, so a scan that says "move me to the system folder and delete the rest" can at most change one allowed label. The official checker passes with zero warnings, and 92 mock-NAS tests cover adversarial model answers.

What we learned

Treating the on-device model as an advisor rather than an actor keeps the convenience of AI sorting while removing prompt-injection risk from file operations.

What's next

Validate on a real iDX6011 Pro: real extract and vision-vlm output, gateway approval prompts and recycle-bin behaviour, using the validation plan in the proposal PDF.

Built With

Share this project:

Updates

Submission history