The problem

Belongings photos and purchase documents already exist in backup folders, but a useful home inventory still takes manual entry. The difficult part is connecting an item to its supporting receipt, knowing when the evidence is uncertain, and maintaining those links after another purchase.

What Home Inventory Ledger proposes

Home Inventory Ledger is an initial UGREEN AI NAS Skill that reconstructs a portable Markdown inventory from authorized photo and receipt folders. It keeps possessions, source facts, receipt line items and user decisions separate. Each relationship shows its evidence and blocker; unreadable or ambiguous records stay visible for review.

A new receipt can resolve an older gap. Rejecting one item/receipt pair does not prevent a different receipt from being considered later. Multiple photos do not silently become multiple possessions or get merged on visual similarity. Existing confirmations and corrections survive updates, while changed source bytes supersede old evidence.

Why it fits AI NAS

The proposed workflow uses the photos and documents already stored on the NAS. Registered file experts discover, hash and retrieve sources; photo experts supply OCR and visual observations; document experts extract printed receipt fields; knowledge experts save versioned ledgers. The parent is orchestration Markdown, not an executable script or background service.

The Skill makes no direct external requests. Expert model routing follows the device configuration; local-only inference is not claimed. Original files and earlier ledger versions are left untouched. Core saves require an unused filename and complete read-back; support for exact safe creation is still a device-validation dependency.

What exists today

The initial Skill and four linked references are authored. The official static checker passes with no errors or warnings, and 32 local integrity tests pass. Fresh Codex evaluators exercised mocked expert inputs, including a code-free update replay that preserved decisions and retired superseded evidence.

Local Apple Vision OCR recovered some independently transcribed device text in five of six licensed public photographs. The washer photo yielded only camera overlays. That failure led to explicit overlay exclusion and stricter model-role handling. These are substitute-engine tests and mocked workflow evaluations, not UGREEN runtime results or a general accuracy benchmark.

What finalists would validate

On the assigned NAS: install the Skill; probe OCR, receipt extraction and result-size limits; run first-inventory and incremental-update cases; verify the complete saved ledger and unchanged source/version hashes; test interrupted saves and ambiguous evidence. The demo should show one supported relationship, a refused guess, and a later receipt closing an existing gap.

Limits and disclosures

The pilot is bounded to twenty unique inputs, twelve active possessions and a 32 KiB ledger, using item-focused photos. No valuation, insurance decision, scheduling, direct external integration or automatic off-device backup is offered. Installation, target-model behavior, permissions and persistence have not been validated on UGREEN hardware.

OpenAI Codex revised the Skill, fixtures, checks and submission materials. Apple Vision was used only for local input tests. Third-party source/license details are included in the proposal and Skill disclosures. Public test photos and fabricated receipt fixtures are excluded from the submitted ZIP and establish no actual purchase or ownership.

Built With

Share this project:

Updates

Submission history