Cheng-Han Lin

Three shipped products, each built solo. My work is about making large language models into something that can be trusted: retrieval-augmented generation so the output is traceable, self-hosted inference so sensitive data stays put, and machine learning for the forecasting an LLM is bad at — plus all the backend, mobile and payment infrastructure underneath.

I care about evaluation method, privacy design, and the failure cases most people skip.

Projects

HeartBox — AI mood-journalling platform (RAG + emotion forecasting)

Solo developer · Capstone · heartbox.tw · source

A mental-health platform built end to end — ML, backend, frontend, mobile and infrastructure. Live at heartbox.tw, sign in with test1 / test1, no registration. Stored entries and charts all work; live AI generation needs the self-hosted inference service, which is not currently exposed — email me to arrange a walkthrough.

• RAG pipeline

  BGE-M3 embeddings over a ChromaDB store built from seven clinical sources (WHO, APA, NHS, NIMH). Retrieval fires when sentiment falls below −0.4, returns the top-3 passages, and the model must cite them — so advice is traceable rather than invented.

• Random Forest forecasting

  53 features (12 metrics × 4 lag windows, plus 5 calendar features). 5-fold CV: MAE 0.22 on sentiment and 1.04 on stress (22,796 rows); AUC 0.948 with 88% recall on high-stress days (31,720 rows), deliberately tuned for recall because a missed warning costs more than a false alarm.

• Model selection

  Compared linear regression, a decision tree, Random Forest, XGBoost and an LSTM on data requirements, interpretability, inference cost and overfitting risk. Random Forest was the only one that trains on a few hundred rows, predicts in under 50 ms on CPU, and explains itself.

• Self-hosted inference

  Open-weight Qwen2.5-7B-Instruct on my own GPU behind FastAPI and a Cloudflare Tunnel, so journal text never leaves a controlled environment.

• Privacy and safety

  Journals are Fernet-encrypted (AES-128-CBC + HMAC-SHA256) with keys held apart from the database, plus 2FA. Three-step separated consent per GDPR Art. 7, where refusing AI-training use costs no functionality, and parental verification for ages 13–17. Crisis-keyword detection surfaces national helplines.

Also delivered: PHQ-9 / GAD-7 assessments and an Android build via Capacitor.

LapseWatch — Subscription renewal reminders

Solo developer · Jun 2026 – present · lapsewatch.smallworks.app

• Notifies before a subscription auto-renews, over LINE, email and desktop, with Google Calendar sync and CSV / JSON export.
• A Chrome extension that detects subscriptions while reading only a narrow window of text around a matched keyword, to bound what it can see.
• Integrated recurring subscription billing and statutory e-invoicing (ECPay, Taiwan’s e-invoice system).
• Privacy-first: no bank-account linking, no cross-site tracking.

PantryKeeper — Shared household inventory tracker

Solo developer · Jun 2026 – present · pantrykeeper.net

• Multi-user inventory and expiry tracking with per-member permissions, offline support and waste statistics.
• A bulk import parser that accepts free-form data pasted from Notion, Excel or Google Sheets and infers quantity and unit from it.

Skills

Education

National Chin-Yi University of Technology

M.S., Computer Science and Information Engineering · From September 2026 · Taichung, Taiwan

National Chin-Yi University of Technology

B.S., Computer Science and Information Engineering · September 2022 – June 2026 · Taichung, Taiwan

Capstone project: HeartBox (above)

Contact

alan930604@gmail.com · github.com/alanlin0604 · linkedin.com/in/chenghanlin-tw