ML engineer — vision-language inference, applied NLP, on-device AI. Taipei, Taiwan · open to remote · available from September 2026
I turn customer requirements into system design, then debug what breaks. Most of my work is production machine learning: getting models onto real hardware behind real APIs, with the operations documentation that lets someone else run them.
Mar 2026 – Apr 2026 · Taiwan · AIoT computer vision for traffic, access control and public-safety deployments
- Gathered requirements directly from the customer and designed the system — replaced a pipeline needing a separately trained YOLO model per scenario with one generalising vision-language deployment, specifying an 18-field structured output schema (visibility, vehicle breakdown, congestion, incident and lane-blockage detection, recommended action) that downstream software could consume without changes.
- Diagnosed two deployment-blocking failures under time pressure — blank
analyses traced to images nested inside
messagesrather than at the/api/generateroot, and degenerate repeating output from parallel small models thatrepeat_penaltycould not suppress, resolved by re-sizing to a single larger model after measuring both configurations. - Delivered the full stack and handed it over — Apache/PHP portal → FastAPI
service → Ollama → MySQL audit log, with
systemdauto-start, an API-key boundary, a health endpoint covering every tier, a spec-compliant REST API for third-party software, and complete operations documentation for another engineer to run it from.
Built as StreamSight.
| Project | What it is |
|---|---|
| StreamSight | Real-time VLM inference platform for IP cameras & IoT — FastAPI + Ollama concurrent pipeline, PHP/MySQL portal, systemd deployment, spec-compliant REST API |
| TechPulse | Offline-first on-device AI learning platform (iOS) — SwiftUI, SwiftData, Foundation Models; every inference local, 68-concept knowledge graph, 49 tests across 7 suites |
| SeedSent | Few-shot LLM-guided ABSA via sequential DeBERTa-v3 fine-tuning — pseudo-labelling that removes most of the annotation budget; ×ばつ speedup via batched GPU inference |
| aspect-sentiment-analysis | DeBERTa-v3 on M-ABSA, evaluated against SemEval-2014, ACL-14 Twitter and hand-annotated reviews |
| NCAV_Lasso | M.S. thesis — sparse vs. dense models in explainable AI: NCAV + Lasso interpretability layer on a from-scratch ResNet-50 |
Spec-first, agent-assisted development: specification and architecture written first, implementation driven through Claude Code, with the design document and test suite as the contract.
ML PyTorch · Transformers · DeBERTa-v3 · Ollama · NVIDIA CUDA · multi-GPU · YOLOv7 · OpenPose Backend Python · FastAPI · asyncio · PHP · MySQL · Docker · systemd · Linux Apple Swift · SwiftUI · SwiftData · Foundation Models
M.S. Computer Science & Information Engineering — dual degree University of Queensland, Australia (Feb 2024 – Dec 2025) Tamkang University, Taiwan (Sep 2021 – Jan 2026) Thesis: Comparative Analysis of Sparse and Dense Models in Explainable AI for Image Processing, supervised by Prof. Tim Miller. Teaching Assistant, Introduction to Artificial Intelligence (Feb–Jun 2022)
B.S. Computer Science & Information Engineering Tamkang University, Taiwan (Sep 2018 – Jun 2021)