Starting in a research lab
Researched medical imaging: hemorrhage localization, coronary calcification, circulating tumor cell segmentation. Ran Ubuntu lab servers and deployed LLMs on A100s, the first time systems themselves caught my interest.
Hi, I'm
李浩廷 · Call me Justin.
I turn real-world needs into systems that keep running. My edge: a medical-imaging research background, hands-on hardware experience, and a habit of checking the numbers.

I started in a research lab, using deep learning to read brain CT and cell microscopy images. An internship at a biotech startup showed me the real gap: models and prototypes are everywhere, but things that run every day are rare. So that became my focus: turning lab needs into systems, wiring up cameras, sensors and GPUs, and finding out why when they break at midnight.
Researched medical imaging: hemorrhage localization, coronary calcification, circulating tumor cell segmentation. Ran Ubuntu lab servers and deployed LLMs on A100s, the first time systems themselves caught my interest.
Helped manage internal systems and IT infrastructure, and handled whatever came up. For the first time, textbook knowledge had to answer to real people.
Research and product, side by side: 4+ full-stack systems with Vue.js and FastAPI, run as Docker services on GCP Compute Engine, watched by Uptime Kuma, YOLO on edge devices. The research went on to Japan and Chicago.
Grew internship prototypes into systems that run on their own, and own the moments they break. I also build the internal tools colleagues use every day: a company AI desktop assistant, a lab records app that replaced a paid SaaS, a Shopify back office, plus local AI inference and a PubMed RAG pipeline.
Next is depth: becoming an engineer who designs, deploys and operates reliable software and AI systems, in Taiwan or abroad. I documented the path in my iThome Ironman series.
The value I bring to a team. Each strength is backed by systems that have run in production, not just tutorials.
From requirements to a working product: APIs, databases, auth and the interfaces people actually use.
Keeping systems healthy after launch: containers, CI, monitoring, alerting and fixing what breaks.
Bringing software into the physical lab: cameras, sensors, edge boards and GPU servers.
A foundation in deep learning for medical images, now applied to agent tools and RAG, with the habit of measuring carefully.
When requirements are fuzzy, I don't wait around or start on a hunch. I research the problem and possible directions myself, lay out a few options, then align with whoever asked so we choose a path together.
For LabMonitor I first trained my own model to read instrument digits. The data was too uniform and it never got reliable. Instead of forcing it, I switched to GPT-4o mini to read the digits and got the system working first.
More than code that only lives in the cloud, I enjoy connecting things to the real world: YOLO on Raspberry Pi and Jetson Nano, LLM and speech models on DGX Spark and Mac mini, A100 lab servers at university.
I care how a number was measured. Benchmarking Relay, I ran 3 warm-ups and 30 interleaved runs and stated plainly that the results are local only. In devops-lab I wrote down the scanner gap I couldn't solve instead of pretending I had.
In university I spent a year as the wind band's student conductor: planning rehearsals, choosing repertoire, working with instructors. Getting dozens of people moving in one direction takes planning and consensus, not orders. I bring that to work.
When I want to understand something, I build one. To manage my own dev accounts I wrote Relay, a cross-platform CLI published on npm. To learn the delivery chain I keep devops-lab: one small app taken through image signing, Kubernetes, Terraform and CI.
From a vendor-independent internship prototype to a second version with cameras, OCR, sensors and about 12 alert types, with no service outage while in operation. Materials cost about NT$20,000 versus an NT$120,000 alternative.
→ Build · Run · Connect
Colleagues needed AI that could work with confidential data. I led the company customization of an open-source desktop agent: company SSO, the internal model gateway, beta/LTS releases with auto-update, a web portal, and 7 agent tool families such as long-audio transcription and approval-gated sharing.
→ Build · Run · Research
Moved the lab's records off Ragic, a paid SaaS, into an in-house app built around its own workflows: reagent lots, per-tube checkout and instrument records, behind company Azure AD. Four subscription seats are already cancelled.
→ Build
Set up and operate LLM, speech and video-generation models, with health checks and auto-recovery from GPU memory errors.
→ Run · Connect
Worked on video management integration: multi-stream RTSP, live browser viewing, recording and clip extraction.
→ Build · Connect
Turned PubMed abstracts from 2016 to 2026 into a semantically searchable knowledge base on PostgreSQL and pgvector, with LangSmith tracing of the RAG flow.
→ Build · Research
A chat-driven meal logger with its own OAuth-protected MCP server: about 20 tools shared by the in-app agent and by Claude, plus a matching Agent Skill. I use it with Claude every day.
→ How I work: still building
A cross-platform CLI for managing Claude Code credential snapshots, with locked writes and atomic file replacement.
→ How I work: still building
One FastAPI app carried through the delivery chain: a signed chiselled Ubuntu image, a hardened Kubernetes deployment that rejects unsigned images, Terraform-managed monitoring, and CI on kind. Runs on a laptop, not production.
→ Run · How I work: clear limits
I enjoy getting on stage and making something complex clear. My master's work on circulating tumor cells went to Japan and the US.
My first international talk: deep learning to locate hemorrhage in stroke patients.
First-author paper on detecting coronary artery calcification with deep learning.
My master's research on stage: automated segmentation of cultured circulating tumor cells.
First-author poster on an automated approach for real-time tumoroid monitoring.
Co-author of a peer-reviewed paper in Diagnostics 15(24).
Cell segmentation, time-series and one-class anomaly detection. Advisor: Dr. Ting-Ying Chien.
Music, snow and good company: the parts of me that don't fit on a resume.
I played trombone in the Yuan Ze University wind band, then spent 2021–2022 as its student conductor: leading rehearsals and performances, choosing repertoire, and preparing us for the National Student Music Competition.
Snowboarding is my favourite way to switch off. For the past three winters I have spent about ten days riding in Hokkaido.
I love joining group activities. It's the same reason I loved the band: lots of people, one shared goal.
I'm in no rush to pin myself to one job title. Over the next three years I want to become an engineer who can design, deploy and operate reliable systems, finding my specialty through hands-on work.
TypeScript work is AI-assisted.
My contest series, from an SRE lab to production AI reliability: FastAPI, Prometheus, Grafana, Loki, OpenTelemetry and LLM tracing, applied to SLOs, error budgets, failure injection and incidents.
Build → Trace → Break → Measure → Evaluate → Recover → Improve
Hiring for backend, internal tools, platform, AI engineering or SRE? Looking for a research collaborator? I'd love to hear from you.