contract-invoice-vault-auditor

• 

  5-Stage Local-First Agent Orchestration Pipeline across file-agent, photo-agent, document-agent, and knowledge-agent on UGOS Pro.

Inspiration

Freelancers, creative studios, and small businesses store PDF contracts, vendor invoices, smartphone receipt photos, and recorded client/supplier voice memos across scattered NAS folders. Reconciling project billing before tax filing or client audits takes hours of manual cross-checking, catches duplicate invoices too late, or forces users to upload confidential contracts and bank receipts to third-party cloud SaaS tools.

What it does

contract-invoice-vault-auditor is a local-first UGREEN AI NAS orchestration Skill built to the Uliya Skill Development Standard v2.0.0 (kimage-planb full main-agent profile). It performs four-way cross-modal reconciliation across documents, receipt photos, meeting audio recordings, and SHA-256 file hashes entirely on-NAS:

1. Stage 1 — Scope Discovery, Cryptographic Inventory & Duplicate Scan (file-agent): Resolves authorized workspace directories (file-workflow get-user-dir), computes sha256sum digests (ug-filesystem), and scans for duplicate invoices (ug-filemgr scan-dedupe) without deleting files.
2. Stage 2 — Visual Receipt & Screenshot OCR/VLM Extraction (photo-agent): Searches NAS photo/screenshot albums (ug-photo search-picture, search-screen) and extracts payment totals, taxes, dates, and wire references via on-device vision-ocr and vision-vlm.
3. Stage 3 — Multi-Document Extraction & Reconciliation (document-agent): Reads page-count and metadata, extracts structured clauses and line items (ug-document extract), and reconciles contract caps against invoice totals and receipt payments (reconcile, qa).
4. Stage 4 — Verbal Change-Order Verification & Audit Ledger Persistence (knowledge-agent): Transcribes meeting recordings (ug-knowledge audio-transcription), correlates verbal fee approvals (collect-evidence), and saves a Markdown Audit & Reconciliation Ledger (save-text) after workflow approval.
5. Stage 5 — Approval-Gated Semantic Tagging & Vault Archiving (file-agent): Strictly after user and gateway/workflow approval (waiting_approval / waiting_user), attaches semantic audit tags (Verified-Match, Variance-Flagged, Duplicate-Candidate) via ug-filemgr and organizes files into structured vault folders via ug-filesystem (mkdir, cpmv) and file-workflow (batch-move).

How we built it

We designed contract-invoice-vault-auditor using the official UGREEN uliya-skill-kit v0.7.3 and verified every capability against capability-catalog.json (v2.0.0). The package includes SKILL.md and four directly linked reference specifications (cli-command-matrix.md, schema-and-output-templates.md, test-cases-and-edge-recovery.md, and security-and-permission-model.md), passing scripts/check-skill.mjs with "status": "static-pass" (0 errors, 0 warnings, 0 info).

Challenges we ran into

Designing a multi-agent workflow that chains four distinct experts (file-agent, photo-agent, document-agent, knowledge-agent) while guaranteeing strict upstream-to-downstream real-result handoffs, explicit empty-result (zero matches) termination, partial-batch failure isolation, and mandatory waiting_approval boundaries before any state-changing operation.

Accomplishments that we're proud of

• Four-way cross-modal verification (PDF contracts/invoices + OCR/VLM receipt photos + transcribed voice memos + SHA-256 deduplication) 100% locally on UGREEN AI NAS.
• Zero-error, zero-warning static-pass on the official UGREEN uliya-skill-kit validator.

What's next for contract-invoice-vault-auditor

In Round 2, we will validate the full 5-stage workflow on the assigned UGREEN AI NAS iDX6011 Pro (64GB) environment across our synthetic multi-modal contract/invoice/receipt/audio test suites and record the 3-minute live runtime demonstration video.

Built With

• document-ai
• linux
• markdown
• multi-agent-orchestration
• ocr
• ugos
• ugreen-ai-nas
• uliya-skill-kit
• vlm