Software engineer focused on backend services, full-stack products, and applied AI. M.S. in Technology Innovation at the University of Washington GIX (2027), with a B.S. in Computer Science and a background in biochemistry.

Most of my recent work sits where AI meets real constraints: exact solvers where heuristics time out, pipelines that have to be cheap enough to run every day, and systems that keep a person in the loop on every AI decision.

Portfolio · LinkedIn · johnhuang105@gmail.com

AI-assisted operations platform used by 150+ students and 140+ mentors across two University of Washington capstone programs. One Next.js app, nine FastAPI services, ~1,800 automated tests.

• Team formation on an exact solver. OR-Tools CP-SAT with a five-stage lexicographic objective that proves the assignment optimal and returns the same result on every machine. Redundant capacity cuts took a hard cohort from a timeout to proven optimal in 0.8 seconds.
• Multi-persona critique panel. One vision pass builds a deck digest, persona agents critique it, and consensus counts and slide references are computed and validated in code so the model can't invent them.
• Monorepo merge. Reunited two codebases that had diverged for months under one rule: share mechanism, never policy. CI checks every service against typed agent contracts.

Python FastAPI SQLModel OR-Tools ChromaDB Next.js TypeScript GitHub Actions

Read the write-up

Secondsite · Software Engineering Intern, Computer Vision Turned live crane-camera footage into georeferenced aerial maps across 11 coordinated Docker services. Benchmarked six AWS configurations to cut cost per map from $11.80 to $2.90, replaced hand-tuned masking with SAM2, and delivered the project's first survey-aligned map at 1 cm per pixel. Read the write-up

T-Mobile · Thoth Project Proof-of-concept system that captures expert knowledge through guided interviews and cites its source on every answer. Placed 1st among competing teams in blind benchmarks on both accuracy and speed (under 8 seconds for 95% of requests) by routing simpler requests to a cheaper model. Read the write-up

InboxIQ AI email triage shipped to production: Gmail sync, background workers, per-user data isolation, and a multi-step LLM pipeline with provider failover and an eval harness holding accuracy at 85% or better. Next.js FastAPI Supabase Vercel Railway Read the write-up

Python · TypeScript · SQL · FastAPI · Next.js / React · PostgreSQL · Docker · AWS (EC2, EBS) · GitHub Actions · OR-Tools · OpenCV · PyTorch