Backend / full-stack engineer — PostgreSQL & time-series data, event-driven reliability, architecture decisions. Taiwan (UTC+8) · async-first · open to remote roles · LinkedIn · Blog · [email protected]

• Multi-tenant fleet platform (Jasslin, 2025–2026). Primary platform owner in a two-person team; about 18,000 registered vehicles reporting at 1 Hz.
  • .NET 8 modular monolith, NATS JetStream, TimescaleDB, PostgreSQL row-level security, and a React dashboard.
  • Representative production chart queries went from 10–60+ s to 1–3 s.
  • Nightly 3–4 hour reports were replaced by incremental computation about 2–3 minutes behind live data.
• Real-time surgical-video inference (Smart Surgery, 2023–2025). NVIDIA Holoscan inference core and operators, integrated with C# via P/Invoke, plus a later C++/TensorRT pipeline delivered as a contract.

Decisions go into ADRs and implementation plans with acceptance criteria before code. AI agents do most of the implementation; I own scope, architecture, tests as acceptance gates, and diff review. Model output is not accepted as evidence by itself.

A test-backed reference for database-to-message reliability: transactional outbox, idempotent inbox, FOR UPDATE SKIP LOCKED workers, retry/dead-letter, and a deterministic red-to-green reproduction of a concurrency race.

Bounded-queue backpressure, semaphore-limited workers, capped backoff, and ID-based dedupe, with race-tested CI. A small demo, not a production system.

• FOR UPDATE SKIP LOCKED Was Not Enough: A Stale Failure Write Race in an Inbox Processor

• VoxProof — Rust transcript-QA prototype; an example of keeping AI-assisted work behind explicit review boundaries.
• Gargantua Relativistic Renderer — Rust research renderer with deterministic, validated artifacts.
• Personal site — no-framework, single-page site with unit and end-to-end tests.