About

I build the part that has to keep running.

The interesting problems show up after the model works: what it costs to serve, what breaks under bursty traffic, and what the agent actually retrieved.

I'm Matthew Hong, an AI engineer in Taichung, Taiwan, with a backend background in Go. Most of my work is the infrastructure around AI systems: the serving path, the agent's memory, and the instrumentation that says what actually happened.

I did not start in software. I studied financial engineering and actuarial science, worked as a financial analyst and an actuarial consultant, then as a data analyst at a precision-manufacturing plant, where I replaced the paper inventory and procurement workflow with the factory's first ERP system. That got me moved into software, and I have stayed there since: backend for a computer-vision safety platform running on cloud GPUs and Jetson edge devices, then Go services with a hexagonal architecture and a full OpenTelemetry, Prometheus and Grafana stack.

Alongside that I finished a master's in computer science at National Chung Hsing University. The thesis benchmarks three ML serving architectures under CPU limits; the second study led to batch2go, an open-source Go proxy for isolating where batching happens. Outside my day job I build taidle, my own product. It started from that job: clients kept asking for AI, and the engineers around me needed something they could call like any third-party API rather than a retrieval stack to run. It is now the customer-support assistant on three client websites, integrated by two independent developers into their own CMS. The engine is private; the design notes and decision records are public in taidle-labs. What I learn on the way ends up on this site rather than in a notebook.

Experience

2026 – nowFounder · taidle (independent)
My own product: a multi-tenant knowledge base with an LLM agent over it, live as the customer-support assistant on three client websites through two developers who integrated it into their own CMS over SSE. Started as a Python/LangChain service for my day job; rebuilt in Go when its evaluations could not tell model variance from system variance. A cassette record-and-replay layer now separates the two, and every reply renders the route it took: what was retrieved, which tools fired, which policies were checked.
2026 – nowAI engineer
Proposed the role and took it on: the first AI engineering function at a consultancy. I scope client AI work across vision and text, starting with whether the data may leave the building and, if not, what hardware it needs, then choose models against latency and experience targets. That gap, clients asking for AI and engineers with no reason to run a vector database, is where taidle started.
2025 – 2026Backend developer
Established Go as the backend standard: a modular hexagonal architecture on uber/fx, isolated schemas per domain module, raw SQL repositories. Built the observability stack end to end — OpenTelemetry Collector into Prometheus and Grafana, Tempo tracing, container and connection-pool metrics.
2024 – 2025Product engineer
Promoted from junior backend engineer within three months. Led technical requirements and delivery across stakeholder and engineering teams for an AI site-safety computer-vision platform; managed cloud GPU and on-site Jetson edge deployments and their cost constraints; instituted a design-first OpenAPI workflow (contract, mock server, then implementation).
2021 – 2023Data analyst → software engineer
Independently designed and delivered a precision-manufacturing plant’s first full ERP in C# (.NET MVC), replacing paper-based inventory and procurement workflows — the work that turned the analyst role into an engineering one.
2019 – 2020Financial analyst · actuarial consultant
Automated financial reporting pipelines in Python, cutting analyst processing time by 82%; actuarial risk modelling and statistical analysis for corporate insurance clients.

Education

2024 – 2026M.S. Computer Science
National Chung Hsing University. Thesis: characterizing ML serving architectures in CPU-constrained environments — the work behind Inference Arena, and the second study that batch2go implements in the open.
2013 – 2018B.S. Financial Engineering & Actuarial Science
Feng Chia University. Founded and chaired the department’s first student committee for three consecutive years, managing its operating budget.
Back to top