Data Lead · Principal Data Engineer.

A content line on a single Mac: picks topics from frontline sources daily, rewrites them in Traditional Chinese, fact-checks, auto-publishes to several social channels, then sediments them into an owned website. 3 channels, ~40k views/month, zero daily human ops.

Engineered for "reads human, never wrong" as a reliability problem — a fact-check gate, a four-layer publish defense, de-AI style rules, and an engagement feedback loop. This site is one of its outputs.

Gemini runs inside BigQuery (AI.GENERATE) to classify comment sentiment and post topics and to write commentary grounded only in the current figures; Vertex embeddings + KMEANS turn negative comments into named themes. A comment RAG layer is gated by a weekly LLM-as-judge evaluation — when retrieval quality drops, the job fails.

A four-tool newsroom agent is live: no free-form SQL, a mandatory self-check after every draft, hard cost caps, and a weekly replay eval. Shared caching keeps AI cost flat as usage grows.

A group digital-revenue dashboard for C-level, a social-performance dashboard for marketing leads and editors, and a tool that typesets monthly revenue tables into print-ready newspaper PDFs — all on Streamlit + Cloud Run with Google sign-in.

A three-tier ACL lets managers add and remove users themselves, with an audit trail and usage analytics, so nobody has to file a ticket to get access.

Took over a Composer + Apps Script revenue pipeline and replaced it with BigQuery External Tables + Scheduled Queries. Same outputs, same freshness, ~300× cheaper to run each month.

The lesson I keep applying: most "pipelines" don't need an orchestrator — they need a query that runs on a schedule.

Facebook, Instagram, Threads, YouTube and GA4 into BigQuery via 34 Cloud Run Jobs on staggered schedules, modelled in dbt with append-only history — every change preserved, not just the latest state.

dbt source freshness, account-level checks and Cloud Monitoring catch the worst class of failure: a job that succeeds but writes nothing.

A side product I built and run alone. It cleans Taiwan's property-transaction records into comparable prices per usable area (實坪), then serves them through a REST API and an MCP server that share one query core. API keys are stored hashed and usage is metered per account.

Found that idle MCP clients reconnecting every few minutes were burning quota and cloud cost, and fixed it by switching to stateless HTTP. 291 tests, CI that fails closed.