Solo individual entrant: Dr. Hernani Costa | Netherlands

1. What does the Skill do?

Invoice–Order Check turns selected NAS invoices and purchase orders into a source-cited table of discrepancies and unresolved evidence for human review.

2. Target user and problem

Independent consultants and small-business owners need to check supplier invoices against orders scattered across their NAS, without manual uploads to a separate comparison tool or mistaking incomplete evidence for a match.

3. Proposed solution

A read-only workflow extracts comparable fields, pairs documents by supplier and order reference, then reconciles currency and amounts on the same billing basis. It separates confirmed differences, supported matches and unresolved checks; it never approves payment.

4. Core user flow

Select and authorize documents and model data handling; confirm fields; extract; reconcile; review a table showing both source names, observed values and supporting excerpts. A coverage summary exposes every unread document and unchecked field.

5. Substantive difference

Instead of a generic summary or unexplained anomaly score, each finding exposes its evidence and comparison basis. Missing totals, ambiguous matches and partial invoices remain unresolved. This gives the reviewer an actionable check list without claiming accounting-system replacement.

6. NAS capabilities and steps

The NAS supplies the authorized document collection and registered document expert. After selection, ug-document extraction structures invoice/order fields; reconciliation compares those same sources. NAS document capabilities perform the work rather than merely hosting an external application.

7. Technical approach

One Markdown Skill delegates exclusively to document-agent through supported ug-document extract/reconcile capabilities. Explicit document selection bounds reads; extraction gates comparison. The platform-configured model reasons over authorized text; prefer local inference if supported. No added API, export, executable or model weights. Unknown model destinations require clarification before reading.

8. Current development status

Implemented: the Markdown workflow, permission/error states, evidence requirements and disclosures. Official kit 0.7.3 static checks pass. Not implemented or verified: NAS installation, runtime behavior, extraction accuracy, latency and model routing. These await finalist access.

9. Finalist validation plan

Core tests: matching synthetic invoice/order yields supported matches; a planted amount/currency discrepancy is correctly source-cited. Edge case: an unreadable total stays unresolved. Score field extraction, comparison correctness and coverage; record latency and confirm unchanged sources. Repeat with embedded payment instructions to verify they remain inert data.

10. Known limitations and dependencies

OCR errors, ambiguous supplier/order identifiers and incompatible tax or billing bases can prevent comparison. No currency conversion, fraud detection or tax advice. Runtime depends on UGOS document-agent, document capabilities and configured model availability; inference locality remains unverified.

---

Third-party, AI-assistance and data-flow disclosure

Sponsor resources and applicable terms

UGREEN authored the UGOS Skill Creation Tool 0.7.3 and capability catalog 2.0.0. They were used to establish supported capabilities and run the bundled static checker. Source: https://ai.ugreen.com/pages/hackathon-starter-kit . These are sponsor-provided development resources governed by the linked UGREEN User Agreement, Privacy Policy and Skill Content Safety and Publishing Guidelines. No separate open-source licence is asserted and none of the kit, checker or catalog is redistributed.

AI assistance and licence

OpenAI Codex was used for drafting, parallel rules/quality review, workflow revision and packaging assistance, using public documentation, synthetic workflow design and entrant-supplied application context. It is a proprietary development service governed by the account’s applicable OpenAI service terms (https://openai.com/policies/). No Codex software, API client, model weights or third-party Skill is bundled.

Other third-party material

No third-party SDK, open-source library, client document, dataset, image or model is incorporated in the Skill ZIP. The PDF was generated with ReportLab, a BSD-licensed document-production library (https://www.reportlab.com/); it is a development tool and is not included in the Skill.

Runtime data flows and permissions

Runtime flow: selected NAS documents → UGOS document-agent/configured model → in-session report. No Skill-added external API, telemetry, export or sharing. Actual model/provider, locality, retention and external text destinations remain unverified until device access. Before any document read, routing must be known and authorized; otherwise the Skill waits for clarification. The ZIP contains no real invoices or personal data.

Built With

• document-agent
• markdown
• openai-codex
• ug-document
• ugos
• uliya