Haorui Yu
AI production QA: release gates, LLM/VLM evaluation, MCP
New to Contra
Haorui is building their profile!
VULCA runtime: from an approved brief to a human-released package. One approved brief goes in; a checked, human-released package with a manifest and delivery receipt comes out. Eight stages: compile the job, adapt first, check independently, repair once, route exceptions, human release, package deterministically, record delivery. This is the scope of my paid pilot. Current evidence tier: LOCAL_TESTED runtime candidate. Not a deployed service and not customer-confirmed; the visual is a fictional demonstration. vulcaart.art
0
4
Two systems I own end to end. VULCA Visual Systems: 21 MCP tools, 13 cultural, design and media profiles, 11 ComfyUI nodes, verified PyPI release 0.23.1, Apache 2.0. Academic Writing Toolkit: 20 Claude Code skills with 112 regression tests for manuscript audit and prose measurement. Both are open engineering artifacts, not customer deliverables.
1
9
Evidence trace: paired candidates, audit log, human review not run. What an inspectable decision record looks like: a stronger-control and a weaker-control candidate for the same brief, a feedback overlay marking what must change, and an audit log excerpt (mode synthetic_demo, review status human_review_not_run, public boundary method_demo_only). Method demonstration only, not customer work.
1
5
Packaging study: tea system, round 2, L3 cultural check (synthetic). Packaged demonstration from the VULCA studio: a tea packaging family generated, checked at L3 (cultural context) and revised once. Shown as a versioned retrospective. Synthetic study, not client work; no generation runs in the browser. Labels follow the evidence rules published on vulcaart.art.
1
3