Chaoran Huang

I build AI-assisted, event-driven systems for regulated industries. I'm drawn to zero-to-one problems — taking an ambiguous idea all the way to a production system that real users depend on — and I care about the boring parts that make that possible: clear service boundaries, reproducibility, and systems that stay predictable under pressure.

• Founding engineer on an AI-assisted plan-review system that embeds models into regulatory review while keeping behavior deterministic, traceable, and safe to roll out. These days that mostly means evaluation — figuring out how you actually know an AI reviewer got it right.
• Before that, took a greenfield utility-billing SaaS (React · GraphQL · Node.js · Kafka · PostgreSQL · AWS) from zero to a production platform serving real customers.
• Thinking a lot about distributed systems, event-driven architecture, applied ML/LLMs, and applied cryptography.
• Background: M.S. Computer Science, Brown University · B.S. Computer Science, UC Irvine.
• Away from the keyboard: film photography and Brazilian Jiu-Jitsu.

Long-form engineering notes at learn.chaoran-huang.com — the kind of explanations I wish I'd had while learning.

• From Definitions to Computing on Secrets — six parts: what "secure" means, a secure channel, authentication, zero-knowledge proofs, and computing on data nobody can see.
• From One-Hot to BERT — a geometry-first history of how NLP learned to represent meaning, from one-hot vectors to contextual embeddings.
• From Representation to Behavior — the roadmap from a pre-trained transformer to an aligned, reasoning assistant (in progress).
• VCG Auctions, Gently — a story-first introduction to welfare, counterfactual payments, and truthful bidding.
• How the Shape of Demand Sets Your Price — a geometry-first walk from a high-school derivative to optimal pricing.
• Bits, Math, Hashes, Streams — a systems refresher on bit ops, modular arithmetic, hashing, and streams.

• Activation Checkpointing & Tensor Swapping — automatic, graph-level memory optimization for DNN training built on torch.fx; cut peak activation memory ~70–85% by profiling the forward/backward graph and choosing what to recompute vs. swap. (Harvard CS265)
• Athlete Chest X-ray Abnormality Detection — YOLOv5 vs. Faster R-CNN for localizing abnormalities in medical imaging, from DICOM preprocessing to evaluation. (Tongji University research, with paper)

• Effect-TS/effect — fixed multipart upload limits being silently ignored (PR #6395, merged), and reported root-caused bugs in the @effect/ai Amazon Bedrock provider (extended thinking, reasoning tokens), both since fixed: #6185, #6186.