Inspiration

Year-end audit prep is mostly a scavenger hunt. The auditor sends a request list of 30 to 80 documents: twelve months of bank statements, specific invoices, leases, asset purchases. The documents usually already exist on the company's NAS, but they're named scan_0042.pdf or IMG_1234.jpg and scattered across folders. Someone opens them one by one, sends the wrong month or the wrong account, and the back-and-forth drags on for weeks. Audit tools like Suralink and DataSnipper help the auditor once files are uploaded. Nothing helps the business get its own messy folder ready before that.

What it does

AuditReady is an AI Skill for UGREEN AI NAS. You give it the auditor's request list and point it at your finance folder. It:

  • Reads every file by its content, not its filename, including scans and phone photos of receipts.
  • Checks each request line for the right entity, period, bank account, amount and document number.
  • Marks every line Found, Mismatch (with the exact reason, such as "account ending 9930, request asks for 4821"), Missing (named exactly, such as "April 2025 statement"), or Needs review.
  • Ties contract amounts to their invoices.
  • Shows you the full plan first. After you approve, it copies the matched files into an evidence pack (one folder per request item), tags the originals, and saves an evidence index, a gap report and a SHA-256 manifest.
  • On a re-run, only reads new or changed files and tells you what changed: "1 gap closed, 0 new mismatches."

Your originals are never moved or changed, and nothing leaves the NAS.

How we built it

AuditReady is a Markdown-only orchestration Skill built with the official UGOS Skill Creation Tool (uliya-skill-kit 0.7.7, catalog 2.2.0). It delegates each step to UGREEN's registered expert agents:

  • document-agent: parses the request list, extracts fields with OCR, and reconciles contracts against invoices
  • photo-agent: OCR for image receipts
  • file-agent: scans the folder, takes SHA-256 checksums, copies files into the pack, adds tags
  • knowledge-agent: saves the index, gap report and manifest

Each step passes only real results from the step before, and stops if a needed value is missing. Every write (copy, tag, save) goes through platform approval. All five delegation states are handled: completed, failed, cancelled, waiting_user, waiting_approval. The Skill passes the official checker with 0 errors and 0 warnings.

Challenges we ran into

  • No NAS for Round 1. We designed against the capability catalog and the checker, so behaviour on real hardware still has to be proven on the finalist test day.
  • File search matches names, not content. Matching by content means reading every file in scope, so we added SHA-256 change detection to make re-runs fast, and a 200-file limit that prompts the user to narrow the scope.
  • Keeping it trustworthy. An audit tool that guesses is worse than useless. We made "Needs review" a first-class status, so unreadable or low-confidence files are never marked Found.
  • Meeting the format. Our first draft failed the official checker on section structure and delegation handling. We rebuilt it to the orchestration contract.

Accomplishments that we're proud of

  • It answers the question a chat assistant can't: "What's missing, and what's wrong?", not just "what exists".
  • Mismatch checks that read like an auditor's review notes, with the reason stated.
  • A manifest that does double duty: fast re-runs for the client, and evidence for the auditor that files weren't changed after the pack was built.
  • A clean pass on the official checker, with full approval and failure handling.

What we learned

  • On a NAS, the AI's value comes from orchestrating capabilities the box already has (OCR, extraction, checksums, tags), not from adding another chat window.
  • Designing "empty" and "uncertain" outcomes carefully matters as much as the happy path.
  • Privacy is a real selling point here: bank statements and payroll are exactly the files businesses don't want to upload anywhere.

What's next for AuditReady

  • Validate on a real UGREEN AI NAS: accuracy on scans, speed on 30 to 200 files, approval pacing.
  • Read the general ledger and fixed-asset listing from spreadsheets for full tie-outs.
  • More request types: payroll, GST filings, loan agreements.
  • The same engine for other checklists: ISO audits, tender submissions, insurance claims and grant applications.

Built With

  • claude
  • document-agent
  • file-agent
  • knowledge-agent
  • markdown
  • ocr
  • photo-agent
  • qwen3
  • sha-256
  • skill-md
  • ugos
  • ugreen-ai-nas
  • uliya
  • uliya-skill-kit
Share this project:

Updates

Submission history