UGREEN Hackathon 2026 — Round 1 Project Proposal
Skill: agent-inbox Track: Automate Your Workflow Entrant: Riandi (solo) — Devpost: Jonathan Wilson (user_id 10858315)
1. What does your Skill do? (one sentence)
Agent Inbox is an autonomous intake desk for UGREEN AI NAS: it watches one Inbox folder, classifies new files, extracts key facts, files them in the right folder, and answers questions about the archive, all on-device, nothing uploaded.
2. Who is your target user, and what is their problem?
Freelancers, creators, and solo operators receive a stream of receipts, invoices, contracts, screenshots, and downloads. Files pile up unsorted, key facts stay buried, and finding one paper later means digging through folders or paying a cloud service to index them.
3. What is your solution?
Four experts run on the NAS: a Router notices new files, a Classifier sorts them, an Extractor pulls out amounts, dates, and parties, and a Librarian files them and writes a daily digest. Uncertain files go to a Review queue, never misfiled; corrections become rules the Skill remembers.
4. What is the main user flow?
The user drops files into Inbox; Agent Inbox triages, files, and logs each one. Each morning brings a one-page digest: what arrived, where filed, and what needs a decision. The user can also ask plain-language questions like "find the March electricity bill" and get the file plus its facts.
5. How is your idea different from existing tools or common AI solutions?
Cloud organizers upload private files to other companies' servers; Agent Inbox never lets a file leave the NAS. One-shot classifiers tag and forget; this Skill is a persistent desk that watches, files, learns, and answers questions. Experts plus a review queue flag uncertain files instead of silently misfiling them.
6. Which UGREEN AI NAS features does your Skill use, and at which step?
File: watches Inbox and moves files into the archive (triage, filing). Document: extracts text from PDFs, scans, screenshots (extraction). Knowledge: indexes the archive and answers plain-language questions (search). System: runs the scheduled digest and enforces permission-scoped access to authorized folders.
7. What is your technical approach?
SKILL.md orchestrates four experts (Router, Classifier, Extractor, Librarian) plus an Auditor checking filings. It uses the NAS file, document, knowledge, and system capabilities from the Skill Developer Package. Inputs: Inbox files, corrections, a rules file. Permissions: read/write inside authorized Inbox and Archive folders only. Model: the NAS built-in AI assistant; no external model or upload. Data stays on the NAS; originals preserved, every action logged.
8. What have you built so far? What is not built yet? What still needs testing on a NAS?
Built: this proposal, the complete SKILL.md with the four-expert workflow, reference docs, and a local structural check. Not built yet: NAS CLI bindings, the scheduled digest, and the correction-learning loop. Needs NAS testing: end-to-end filing on UGOS, extraction accuracy on real scans, and Review queue behaviour.
9. How will you test your Skill if you become a Finalist?
Core 1: 20 mixed files (receipts, invoices, screenshots, PDFs) into Inbox; each classified, filed correctly, listed in the digest. Core 2: ask "find the March electricity bill"; the Skill must return the file with its amount. Edge: a corrupted PDF and an unknown file type land in Review with a reason, originals untouched.
10. What are the known limits, risks, or outside dependencies of your Skill?
Extraction depends on the NAS document capability, especially phone-photo OCR; uncertain results go to Review. English first; other languages may reduce accuracy. No external service needed; an external model means that provider's terms apply. The package holds no third-party code, only original Markdown, and runs on the UGOS AI assistant from UGREEN.
Third-party components and licenses: none — the package contains only original Markdown written for this Hackathon; it runs on the UGOS AI assistant and NAS capabilities supplied by UGREEN. AI tools used: drafting assistance from Muse (Meta) for the Skill definition and this proposal; workflow design and all decisions are the entrant's. Submitted for UGREEN Hackathon 2026, Round 1.
