Inspiration
Small filming teams must return cameras and accessories together. The rental manifest and the photos taken when receiving and packing the kit may be scattered across NAS folders. We chose a focused renter-side workflow: show what the selected evidence supports and what a person still needs to check before sealing the box. Customer frequency, willingness to pay and time savings remain unvalidated.
What it does
Gear Return Check compares one user-selected rental manifest with receipt-time images and current packing images. Each listed item receives PHOTO_SUPPORTED, IDENTITY_REVIEW, CHECK_BY_HAND or UNASSESSED, a source reference and a specific next check.
Receipt images cannot satisfy current packing evidence. Repeated closeups of one battery cannot silently count as two batteries. A hidden charger prompts a clearer photo or a manual inspection, never a claim of loss. Readable serial disagreements remain visible. The Skill does not determine damage, functionality, liability or completed physical return.
How we built it
The newly authored Markdown-only Skill follows UGREEN's official Skill Creation Tool 0.7.3 and catalog 2.0.0. It delegates bounded file discovery and fingerprinting to file-agent, manifest extraction to document-agent, image reading to photo-agent and optional reviewed saving to knowledge-agent.
The user authorizes source scope and processing and confirms the governing manifest and image roles. The checklist appears in chat first. Saving requires separate content and destination approval, supported create-only behavior and exact readback. If safe saving is unavailable, the chat checklist remains useful. Originals are not changed or shared by the Skill.
Challenges and lessons
The difficult part is deciding what an observation proves. Similar models are not necessarily the same physical item. A repeated image is not a new packing observation. An unreadable label is not a confirmed mismatch. The workflow preserves these distinctions instead of filling gaps with guesses.
Current status and validation
The Skill and initial ZIP are authored. The original official checker returned zero errors and zero warnings through a read-only, in-memory ZIP adapter; the native-directory CLI was not run. A separate deterministic development reference model passed 28 author-created cases. Those tests do not execute SKILL.md, Uliya, OCR or NAS experts. NAS installation, visual accuracy, actual model routing and save behavior remain unverified. Synthetic development fixtures and the finalist test plan are ready.
What's next
On assigned hardware, test a missing charger followed by a confirming image; same-model cameras with conflicting serials and repeated battery views; and unreadable evidence, reused receipt images and refused saving. Measure row and reference accuracy, unsupported findings, preview time and actual approvals.
AI and third-party disclosure
OpenAI Codex assisted research, capability inspection, Skill authoring, testing tools, static verification and proposal preparation. The entrant reviews and takes responsibility for the submission. The Skill and core workflow were newly authored for this hackathon. UGREEN's organizer kit is a development resource under its applicable terms and is not bundled. Runtime dependencies are registered UGOS experts; no extra runtime API, executable, model weights, credentials or real customer records are bundled. Actual model routing must be established before private-input processing.
Log in or sign up for Devpost to join the conversation.