Durham, NC —
Open to SWE roles · summer + new-grad ’26
Michael Lei · Duke ECE + CS · May 2028 Building reliable systems at the intersection of AI and data.
About
Michael Lei — I'm passionate about turning hard engineering problems into solutions that work in the real world.
I build reliable systems at the intersection of AI and data, and I care about the parts that have to hold up under real use: correctness, scale, and the infrastructure that keeps software dependable in production.
// AI // Data // Systems
Experience
Software Engineer · Durham, NC
Co-founder & Engineer · Durham, NC
Software Engineering Intern · Sunnyvale, CA
Machine Learning Intern · Remote
Software Engineer · Durham, NC
Work
A public CVE intelligence platform on Postgres + pgvector. 340K+ vulnerabilities in one corpus, natural-language hybrid search exposed to LLMs over MCP, and an attack→CVE RAG pipeline strict enough that a hallucinated CVE ID can't happen.
Hybrid semantic search and an attack→CVE RAG pipeline over 340K+ CVEs in one Postgres + pgvector store. The schema makes a hallucinated CVE ID impossible.
Unifying firewall block telemetry across multiple platforms at Duke. AI-driven persistence ranking + coordinated-campaign detection.
30K+ enriched attacks/day into Elasticsearch, false positives cut ~90%, shared across 80+ institutions.
A nutrition + workout tracker for Duke students. LLM agents over the dining database. 3 months from grant to 100+ daily users.
Nutrition + workout tracker for Duke students — LLM agents over the dining DB. 100+ daily users.
An AI code reviewer that takes the first pass on PRs. In production on 3 Bitbucket repos at VSee, including the main product repo.
A composable de-identification pipeline for clinical notes. Regex + ML + LLM detectors stacked end-to-end, configured as data.
Part of cutting clinician charting from ~15 min to ~1 min (~93%) in a customer demo.
An MCP server that lets non-technical teams query Elasticsearch in English. Nine hardcoded tools beat one flexible-LLM-writes-ESQL tool.
A multi-tenant B2B2C wellness SaaS that grew out of DukeDine — each university gets its own branded subdomain, an isolated data plane, and a dining adapter on a shared engine, with a MealMark AI assistant reasoning over food, training, and weight data. Piloting at 10+ schools.
A multi-tenant B2B2C wellness SaaS spun out of DukeDine — Duke is the first tenant; piloting at 10+ universities.
A self-hosted service that predicts ICD/CPT codes from clinical notes and returns token-level evidence spans for every prediction.
Token-level evidence spans over 8,000+ ICD codes. The other half of the ~93% charting-time cut.
An iOS + DJI Mavic Pro app that runs a CoreML detector on-device, flagging wildfires and pre-fire hazards in real time.
~95% on-device detection accuracy; U.S. Division Semi-Finalist, 2022 China–US Young Maker Competition.
More
VSee's company-wide internal LLM platform. Open WebUI + MCPO + 8 MCP servers, BAA-compliant, on-prem. 80+ engineers.
A speech-to-text benchmarking platform with a medical-aware WER metric. Drove VSee's migration off self-hosted Whisper.
A coarse-to-fine auto-tuner for a 7-knob computer-vision pipeline. Replaced hand-tuning a slider panel for every assay slide.
Per-pole controllers that raise and lower Duke's flags autonomously. The proxy was decaying — time wasn't flag position.
Writing
Contact
Open to SWE roles · summer + new-grad ’26 — graduating May 2028. The fastest way to reach me is email.