Inspiration

When a dishwasher or a laptop breaks, the proof that would make the repair free usually already exists: a photo of the receipt in a phone backup, a scanned invoice called scan_0012.pdf, a manual whose warranty clause hides on page 37. People pay for covered repairs because finding that proof takes hours. A NAS already holds all of it, privately, in one place.

What it does

Warranty Claim Pack is a Uliya Skill for the UGREEN AI NAS. The user describes the broken item ("the dishwasher from last March"). The Skill delegates to the NAS photo, file, document and knowledge experts to find receipt photos (photo text index plus OCR), invoices and manuals (name and content search, deduplicated), extracts dates, model and serial, quotes the warranty term with file and page, and answers covered, expired or undetermined, with days left. It drafts a claim pack with every source linked and fingerprinted (SHA-256). Saving the pack and tagging the evidence only happen after the user approves. A sweep mode lists warranties ending soon.

Synthetic example from our answer key: on the fixed test date 2026-11-21 the answer is covered, the warranty ends 2027-05-10, 170 days left, purchase on invoice-ars.pdf page 1, term quoted from manual.pdf page 3. The same invoice in pesos and in dollars counts as one purchase, not two.

How it differs

Warranty apps need purchases registered upfront; this works afterwards, on what is already on the NAS. Photo search finds receipts, not warranty clauses or days left, and cloud search needs the photos on vendor servers. A general AI assistant may assume a typical warranty term; this Skill cites file and page and never invents one. The kit's contract-invoice-audit compares documents you hand it; this one finds them first.

How I built it

With the official Uliya Skill Developer Package and capability catalog 2.0.0: SKILL.md plus five Markdown references, checked with the official static checker (static-pass, 0 errors, 0 warnings). 15 of 17 actions are read-only; the parent only delegates, never calls the CLI directly; document text is treated as data, never as instructions. Searches use the product type in the user's language and in English (lavavajillas and dishwasher). All five delegation states are handled: a failed expert is reported while the other branches keep their results, a cancellation stops only the steps that depend on it, a question pauses only the steps that need the answer, and approval needs an explicit yes, so a rejected save writes nothing.

Challenges

Ambiguous dates (03/04), faded receipts and manuals without a clear term: the Skill never invents a term and asks only blocking questions.

What I learned

The hard part is not search but deciding which file actually proves the purchase. A quotation, a cancelled order or a receipt for an accessory of the same model look like proof and are not, and an invoice never says how long the warranty lasts: without the manual, the honest answer is undetermined.

What's next

Round 2: run it on a real UGREEN NAS against synthetic fixtures with a fixed answer key (two core cases and an injection edge case). A second synthetic set is ready for that run: 15 files, 10 of them look-alikes such as a quotation, a cancelled order and an accessory receipt, plus one purchase recorded four ways (invoices in two currencies, a faded receipt and an order confirmation) and a 60-day sweep that must list one warranty once and skip the other.

Built With

Share this project:

Updates

Submission history