Inspiration

Expatriates, migrants, and multinational families hold passports, permits, insurance, and tax papers in several languages and calendar systems. Files are scattered, deadlines are missed, and sensitive papers should not be uploaded to general-purpose cloud AI tools. We wanted to turn the UGREEN NAS from a static file warehouse into a proactive "Cognitive Layer" that safeguards life documents locally.

What it does

Personal Data Concierge is a private, multilingual AI Skill with five modes: Find and Read, Life Admin Radar, Completeness Check, Cleanup Proposal, and Package. It finds documents using multilingual keywords, extracts text via OCR, and builds a deterministic risk-ranked timeline of expiry and renewal dates.

How we built it

The package is a Markdown-only orchestration layer (SKILL.md + references/global-data.md). All NAS work is performed using the delegate command to interact with file-agent (ug-filesystem) and document-agent (ug-document). It strictly avoids direct cli.invoke calls.

Safeguards include:

  • Deterministic Risk Score: Risk = urgency + impact (0-100).
  • Multi-Calendar Conversion: Converts Buddhist Era, Japanese era, and Minguo by exact rule. Unconverted calendars (e.g., Hijri) are flagged, never guessed.
  • Data Minimization: Identifiers are masked to the last four characters.
  • Strict Approval Flow: Mutating actions (Cleanup, Package) require platform-level approval. The Skill never deletes files. ## Challenges we ran into Handling ambiguous dates (e.g., 04/05/2027) and language variants without hallucinating. We solved this by enforcing strict anti-hallucination rules: if a calendar cannot be deterministically converted, it is excluded from the risk ranking until the user confirms. ## Accomplishments that we're proud of
  • Building a complete, production-ready SKILL.md architecture that passed the controlled static checker with 0 errors and 0 warnings.
  • Designing the "Life Admin Radar" — a feature that gives users immediate, actionable insights on what document is expiring soonest.
  • Handling complex edge cases, such as exact duplicate detection using SHA-256 checksums and mapping physical document requirements (PHYSICAL / USER TO CONFIRM). ## What we learned We learned the importance of bounded privacy handling and strict delegation orchestration. AI models should not guess missing data; they must ask for clarification or mark data as UNCERTAIN. ## What's next for Personal Data Concierge
  • NAS Testing: Validating OCR accuracy on identity documents on real hardware.
  • Runtime Date Injection: Ensuring the platform supplies a reliable current date for the Risk Score classification.
  • Expansion: Adding more jurisdiction-specific procedure templates to the global reference data.

Built With

  • automation
  • claude
  • deepseek
  • githubcopilot
  • markdown
  • ocr
  • powercell
  • ugos
  • ugreennas
  • vscode
Share this project:

Updates

Submission history