About

I’m Cheng-Han Lin, based in Taichung, Taiwan. I have built and shipped three products on my own.

I graduated with a B.S. in Computer Science and Information Engineering from National Chin-Yi University of Technology in June 2026, and started the M.S. programme in the same department that September.

What I work on

My work keeps returning to one problem: how do you turn a large language model into something people can trust?

The model itself is the easy part to obtain. The hard part is everything around it — making sure it cannot invent an answer, keeping sensitive data inside a controlled environment, producing a useful forecast from only a few hundred rows per user, and deciding how the system should behave when it is wrong.

Concretely, I have built:

Retrieval-augmented generation — vector retrieval over BGE-M3 embeddings in ChromaDB, where the model is required to cite the passage it drew from rather than answer freely.
Self-hosted inference — open-weight models running on my own GPU service behind FastAPI, so highly sensitive text never leaves an environment I control.
Machine-learning forecasting — feature engineering, model selection and cross-validation, traded off against interpretability, inference cost and overfitting risk.
Backend and infrastructure — Django REST Framework, PostgreSQL, Cloud Run, Cloudflare, and mobile packaging.
Billing and compliance — a complete subscription billing flow including statutory e-invoicing (ECPay, Taiwan’s e-invoice system).

What I care about

I care less about which technology was used than about why it was chosen.

Every decision costs something. Choosing Random Forest gives up whatever temporal structure a sequence model might have found. Tuning for recall — catching as many real cases as possible — means accepting more false alarms. Self-hosting a model means owning its operations. The difference between engineers shows up mostly in whether they can articulate those trade-offs.

The HeartBox case study is written to that standard.

Contact

Cheng-Han Lin standing against a white wall in a white t-shirt and loose jeans, a stack of magazines on the floor to one side