Theme: Automate Your Workflow. A Skill for the UGREEN AI NAS, built for the hackathon. The Round 1 proposal (PDF) and the initial package (ZIP with SKILL.md and references/) are attached to this submission.
One sentence: What does the Skill do?
Monthly, it matches supplier orders, delivery notes and invoices on the NAS and lists price, quantity, duplicate and VAT differences needing action, changing nothing without approval.
Who is the target user and problem?
Small restaurants and shops with many suppliers. In my family's restaurant in Asturias nobody compares each invoice with what arrived, so price rises, missing goods, duplicates and wrong VAT pass unnoticed.
What is your proposed solution?
It reads the month's documents in place, phone photos included, and matches order, delivery note and invoice line by line per supplier, showing both values and the euro gap. A ledger on the NAS carries prices into next month.
What is the Core User Flow?
The user gives folder, month and tolerance, and confirms the supplier grouping. The Skill checks each supplier; the user confirms or dismisses each finding. With approval it saves report and ledger, then copies, renames and tags files in a monthly folder.
What substantive difference does your idea offer compared with existing tools or conventional AI solutions?
A chat assistant needs uploads, forgets last month and cannot organise folders. Bookkeeping software records what an invoice says, not what arrived. This works in place, remembers prices and dismissed findings, and ends in approved folder changes.
Which UGREEN AI NAS capabilities will be used, and at which step in the workflow?
Inventory: file listing, search, hashing, duplicate scan. History: saved ledger read back. Photos: text recognition. Checking: document extraction, reconciliation. Saving: text saving. Organising: folder creation, copy, rename, tags. Documents and history stay in the user's storage; the Skill uploads nothing.
What was your technical approach?
SKILL.md delegates in order to file-agent (ug-filesystem, ug-filemgr: inventory, approved copy, rename, tag), knowledge-agent (ug-knowledge: collect-evidence, save-text), photo-agent (ug-photo: vision-ocr) and document-agent (ug-document: extract, reconcile; one supplier at a time). Inputs: folder, month, archive, tolerance. Reads need no approval; each write needs the user's approval, then the platform's. Instructions are self-contained; all five delegation states are handled.
What is the current development status?
Done: initial package passing the kit's static checker. The idea comes from an earlier tool of mine; nothing is reused. Not done: any NAS run, real-document test, user feedback. To validate: photo text recognition, delivery-note extraction, reconciliation.
What is your final validation plan?
Case 1: synthetic supplier month, four planted errors: price up 8 percent, 12 units invoiced for 10 delivered, duplicate invoice, wrong VAT rate. All must be found, no false alarms. Case 2: next month with the saved ledger: new rise caught, dismissed finding not repeated, copies as approved. Edge: handwritten photo, text-less PDF, invoice without delivery note: reported unreadable or unmatched, nothing invented.
Are there any known limitations of your Project?
Handwriting and text-less scans may be unreadable. The catalog has no delivery-note schema, so extraction quality is unknown. Findings are candidates; the Skill does not decide legal VAT rates. No scheduling. Where the experts' models run is unstated, so I claim only storage. Depends on the NAS experts.
Third-party components and licences
None. No third-party code, libraries, models or datasets are included in the package. I followed the Uliya Skill Development Standard (kit 0.7.3) and ran its static checker locally; none of the kit is bundled. The Skill sends no data to any external service.
Built With
- markdown
- ugos
- ugreen-ai-nas
- uliya-skill-kit
Log in or sign up for Devpost to join the conversation.