Inspiration
What it does
How we built it
Challenges we ran into
Accomplishments that we're proud of
What we learned
What's next for DocDrift for NAS (docdrift-nas)
Entrant: Hwemo Chung, individual entrant, no team. Track: Automate Your Workflow. Contest: UGREEN Hackathon 2026 on AI NAS.
One sentence: What does the Skill do?
docdrift-nas finds a newer source version on a NAS, flags stale clauses in dependent SOPs, contracts, and manuals with quotes and severity, and, after confirmation, saves a redline report per affected document.
Who is the target user and problem?
Target users: operations and compliance teams with NAS documents. When laws, price lists, specifications, or vendor terms change, dependent SOPs and contracts keep old facts, and manual checks are slow, error-prone.
What is your proposed solution?
DocDrift is a Markdown-only Skill using three NAS experts: find source versions and dependents, extract the change set, reconcile documents, show stale clauses with quotes, facts, severity, and rewording. Nothing is written before confirmation; each confirmed document gets a .redline.md report.
What is the Core User Flow?
The user names a source document and dependent folder. file-agent finds baseline and newer versions plus dependents. document-agent extracts changes and returns stale clauses with quotes, facts, severity, and rewording. The Skill asks for explicit confirmation. file-agent verifies a free path; knowledge-agent saves a .redline.md per confirmed document. Originals are never edited, moved, overwritten, or deleted; empty results stop.
What substantive difference does your idea offer compared with existing tools or conventional AI solutions?
Chat cannot prove what it read; search finds words, not dependencies. DocDrift compares two named source versions, tests each dependent clause against the change set, and refuses findings without a verbatim quote and location. New reports only.
Which UGREEN AI NAS capabilities will be used, and at which step in the workflow?
- file-agent (ug-filesystem find, stat): Stage 1 locates versions and dependents; Stage 5 verifies paths.
- document-agent (ug-document extract, qa, reconcile): Stage 2 builds the change set; Stage 3 finds stale clauses.
- knowledge-agent (ug-knowledge save-text): Stage 6 writes a redline .md per confirmed document.
What was your technical approach?
Stages: locate, extract, reconcile, confirm, verify, save. Each uses only verified upstream results and stops on gaps. It delegates one instruction per stage to an expert agent. Inputs: source title and folder, dependent folder, patterns, version roles. Permissions: read-only until confirmation, then one approval-gated save-text per document. The Skill selects no model; data stays on the NAS.
What is the current development status?
Implemented: SKILL.md and two reference files. Passes the official checker (uliya-skill-kit 0.7.3, catalog 2.0.0) with 0 errors and 0 warnings. Not yet implemented: tuning against real NAS documents. Pending NAS validation: all six stages, save-text approval, and report path checks. Never run on a NAS.
What is your final validation plan?
Test 1: price list v1/v2 with one changed fee, two SOPs repeating it; both flagged High with quotes and redlines saved after confirmation. Test 2: v2 removes an obligation, renames an entity; the checklist is flagged with correct categories; an unaffected contract yields nothing. Edge case: one source version only; the Skill reports "No newer source version was found" and writes nothing.
Are there any known limitations of your Project?
Accuracy depends on document-agent's reading; scanned, low-quality PDFs and tables may be misread. Coverage is limited to readable documents, one source family per run, no watching or notifications. Not legal advice; approval needs NAS confirmation. Paraphrases may be missed or over-flagged.
Third-party components and licenses
UGOS Skill Creation Tool / uliya-skill-kit v0.7.3 from UGREEN's Hackathon Starter Kit was used only for the static check, under UGREEN's user agreement. No other third-party code is included. The Skill contains only Markdown: no scripts, binaries, external services, or data flows outside the NAS.
AI tools disclosure
Drafting was assisted by Grok Bot and OpenCode with the DeepSeek V4.1 Flash model. The author reviewed the text and takes responsibility for it. The Skill's own runtime AI is whatever the NAS experts use; the Skill selects no model.
Log in or sign up for Devpost to join the conversation.