Inspiration

Similar product variants can be easy to mix up when preparing listing photos. A cream cup beside a blue backdrop does not necessarily have a blue rim. FoundHere focuses on that selection boundary.

What it does

The owner provides two positive examples and two confusing near-misses, then approves a short visual rule. Predictions on four unseen photos are recorded before the owner reveals their labels.

The rule is then applied to a new batch. Required traits must belong to the same designated object. Hidden or ambiguous details remain review items. The owner selects the originals to copy into a working folder.

How it was built

FoundHere is a six-file declarative Uliya Skill created using the official UGREEN Skill Developer Package. It coordinates file-agent and photo-agent for source checks, visual observations, approved persistence and verified working copies.

Saved rules can be reused after checking their sources and available model/route provenance. Image processing requires a verified local route.

Challenges and learning

Visual uncertainty, cross-object attribute mixing and changed source files can undermine a selection. The workflow explicitly handles these cases. Four check images are an onboarding check, not a statistical accuracy study.

Current status and next steps

The initial Skill passes the official static checker with zero errors and warnings. Local policy checks exercise decision and handoff rules. NAS execution, recognition accuracy and user-time benefit remain unvalidated.

Finalist testing will compare the workflow with native tools, an equally informed generic Uliya request and manual selection.

AI assistance

OpenAI ChatGPT/Codex assisted research, refinement, authoring and local checks. Component sources and licenses are disclosed in the proposal and linked Skill references.

Built With

Share this project:

Updates

Submission history