Inspiration
My family runs a screw factory, where shift handover is verbal and out-of-tolerance lots are often found late.
What it does
At the end of a shift, a supervisor asks Uliya for a brief of one line and one time window. The Skill:
- finds only that shift's files in the line's NAS folder;
- reads the part tolerance sheet and every checksheet value (PDF or photo), and flags values outside the limits with the exact number, limits and source;
- looks at QC sample photos for visible signs such as cracked heads, burrs or damaged threads, and says "cannot judge" when a photo is blurred;
- summarizes machine logs into good/reject counts and stoppages, and attaches the operator voice memo (and, optionally, camera events) recorded at the time of each stoppage;
- writes a brief that marks lots as review or watch — never pass or fail — and lists everything it could not check;
- saves the brief and tags flagged photos only after the user approves.
How we built it
- Read the Uliya Skill Kit and capability catalog, and mapped every step to a published CLI and its single registered expert (file-agent, document-agent, photo-agent, knowledge-agent, security-monitor-agent).
- Wrote the SKILL.md orchestration with self-contained delegate instructions, explicit handling of completed / failed / cancelled / waiting_user / waiting_approval, empty-result stops and a raw-text fallback for unsupported log formats.
- Built a local simulator (Python, not part of the package) that generates a synthetic cold-forming shift — three machines, two lots, checksheet images, sample photos, voice memos — and runs the same workflow against mock experts, with 13 tests for the rules. The package passes the kit's static checker with zero warnings.
Challenges we ran into
Skills can only be Markdown, so all analysis is orchestrated through UGOS built-in experts; we added a raw-text fallback for unsupported log formats and strict no-guessing rules for numbers.
Accomplishments that we're proud of
A Skill package that passes the official kit checker with zero errors and zero warnings, plus a local simulator with 13 tests.
What we learned
How to design reliable LLM orchestration with explicit status handling and evidence links.
What's next
- Validate OCR on handwritten checksheets and Mandarin/Taiwanese memos on a real AI NAS.
- Weekly trend brief across shifts (repeated alarm codes, die-change frequency).
- Pilot at a family-run fastener plant using its own NAS.
AI tools used: Claude Code (Anthropic) and Codex (OpenAI) were used for reading docs, building the simulator and drafting text.
Log in or sign up for Devpost to join the conversation.