MYWU Labs™ · Applied AI · Houston, TX
I’m Michael Wu. I build retrieval systems, evaluation loops, and self-hosted AI stacks that still make sense after the demo — scored against the expert rubric, not the vibes.
Move your pointer inside the panel — the cat hunts it, and every catch pulls a chunk from the corpus. It is the whole job in one picture: the model is not the trick; the retrieval is.
01 — Services
Every engagement starts the same way: with the rubric. What does correct look like, who decides, and how would anyone know the difference between a good answer and a confident one?
RAG pipelines over your own documents — ingested, chunked, embedded, reranked, and grounded so every claim points at a source you can open.
Can an AI grade the way your best human grader does? I build the pipelines that answer that empirically, then tune them until they agree.
Self-hosted AI on your hardware or a single small VPS: Docker, nginx, LiteLLM, SSO, and GPU or Apple Silicon inference. No surprise egress.
LMS and developer-platform engineering — the unglamorous plumbing that makes a course or a cohort run without an IT ticket per student.
Curriculum and faculty enablement for AI — built from twenty years of standing in front of the room, not from a prompt-engineering slide deck.
Independent review before you commit: architecture, code, security, and the honest question of whether the AI is adding value or adding risk.
What is actually breaking, what it costs you now, and what a good outcome would have to prove. If it is not a fit, I say so on the call.
A short spec with milestones, a defined eval set, and a price. Weekly review against your own examples rather than a demo I control.
A runbook, the eval suite, and someone on your team who can run it. The goal is to become unnecessary, which is why people come back.
02 — Selected work
A practice is only as good as the things it runs in production. These are live systems and published evaluations, not proposals.
A self-hosted lab platform serving hands-on exercises across two course tracks: one Keycloak login, browser code workspaces, AI tutoring over code without exposing it, and a routing layer that keeps provider keys server-side. Deployed incrementally with AI agents — every change reviewed by hand.
Can an agentic AI code-evaluation pipeline replicate expert instructor feedback? Built on rubrics developed over years of grading student code, with AI output scored against human review until it matched.
Undavall, T., & Wu, M.Y. — UHCL Student Conference, April 2026.
A feedback system that reads student code and responds the way an instructor would — the same loop as RLHF, applied to grading: generate, score against expert judgment, iterate until the gap closes.
SWBATC, Texas Southern University, February 2024.
An inline chat layer inside course content where each tutor answers only from its own assignment context — slides, coverage maps, and instructions — cites the slide it used, and refuses to hand over the graded answer.
Retrieval is not a feature of the tutor. It is the tutor.
03 — Expertise
Twenty years across systems programming, enterprise backends, web platforms, and statistics — which is why the AI work is grounded in engineering rather than demo craft.
“A confident answer is not a correct one. The work is building the rubric that can tell them apart — and then having the discipline to apply it to your own system first.”
04 — Record
Selected publications and presentations across AI evaluation, computational thinking, statistical modelling, and the systems work underneath them.
Marquis Who’s Who (2025) · IEEE member · Microsoft Office Specialist, Excel 2016 (77-727) · English and Mandarin.
05 — Teaching
Every term since 2015 at UH–Clear Lake, plus graduate sections before that at NJIT. Teaching is where the rubrics come from — hundreds of submissions a term is a much harder test than a benchmark.
06 — Contact
The most useful first message is short: what you are trying to make reliable, what you have already tried, and what would count as a win. Scope calls are 30 minutes and free — and if it is not a fit, I will point you at someone better.