ML infrastructure engineer

Yang-Che Tseng (Jeff)

I build the infrastructure that puts models to work: LLM apps and agents, data pipelines that run in production, and the GPU fleet behind 110k+ Slurm jobs for my M.Sc. in computer science at UBC.

About

I take AI systems from a rough idea to something a team relies on: connecting them to the systems and data that already exist, working with the people who use them, and keeping them running. At GoGoX I built an LLM retrieval assistant and put it in front of beta users, and turned a routing engine buried in a legacy Java system into a service. At MyProGuide I built a serverless AWS pipeline that reads cancellation emails from 12 travel agencies and updates orders; it has run in production for over a year at up to 35k emails a month, and I worked with the operations team to recover 3,200 of 3,600 historical emails.

On the LLM side I build retrieval (OpenAI, LangChain, Pinecone, Redis Stack), text-to-SQL agents and agentic tooling. For my M.Sc. at UBC I built a Claude Code setup with safety hooks that submits, monitors and fixes Slurm jobs across three clusters: 110k+ jobs and about 15.8k GPU-hours in all.

Focus
  • LLM infrastructure & agents
  • ML infrastructure & HPC
  • Prototype to production
  • Systems integration
  • Data pipelines
  • 3D computer vision
Tools
  • Python
  • C/C++
  • CUDA
  • TypeScript · Node.js
  • PyTorch
  • Slurm · Apptainer
  • AWS
  • Google Cloud
  • Cloudflare Workers
  • GitHub Actions
Education

M.Sc., Computer Science · The University of British Columbia · 2024–2026

B.Sc., Computer Science and Information Engineering · National Taiwan University · 2019–2023

Experience

Industry

  1. 2024–2025

    Software Engineer (part-time) · MyProGuide (TravoAI)

    Remote

    • Serverless email pipeline: built an AWS SES → Lambda → S3 pipeline that parses cancellation emails from 12 online travel agencies (Klook, KKday, Viator and others) and updates orders through the API. In production for over a year at up to 35k emails a month.
    • Wrote multilingual per-agency parsers with a Jest suite built on real emails; recovered 3,200 of 3,600 historical emails with the operations team.
    • Flutter app: built a tour-guide app from scratch to MVP through spec-first issues and reviewed PRs (BLoC, OAuth, CI).
    • LLM tooling: prototyped a text-to-SQL agent (LangChain, Gemini) that creates PostgreSQL test accounts.
  2. 2022–2024

    Software Engineer, Data Scientist · GoGoX

    Data Scientist Intern from Sep 2022, full-time from Jun 2023 · Taipei, remote from Vancouver

    • LLM QA bot: built a RAG chatbot on OpenAI, LangChain and Pinecone, deployed on Google Cloud with a Streamlit MVP for beta users; a self-hosted Redis Stack vector store cut 2 s of latency.
    • Routing service: extracted the GraphHopper routing engine from a deprecated Java dispatch system into a Docker Compose service with a Python interface, used for counterfactual analysis of driver stop probability.
    • Labeling automation: an ML model on GPS and order data that labels pickup times automatically, with autoencoder anomaly filtering.
    • GPS map matching: HMM matching to OpenStreetMap (Python, shapely) to build a travel-time database for ETAs.
  3. 2022

    Software Engineering Intern · ASUS Intelligent Cloud Services

    Taipei

    • Extended Selenium IDE with Electron and Node.js, making 19 of 23 previously unrecordable test cases automatable.

Research engineering

  1. 2023–2026

    M.Sc. thesis research, 3D computer vision · UBC

    Research Assistant, Computer Vision Lab, 2023–2024; M.Sc. 2024–2026 · Vancouver

    • Evaluation protocol: found that single-image novel view synthesis benchmarks score misaligned predictions; proposed a fix (matching → PnP → co-visibility mask) and re-benchmarked 11 baselines on 4 datasets (1.7k test views).
    • GPU fleet at scale: ran 110k+ Slurm jobs (about 15.8k GPU-hours) on 3 clusters (A5000/3090, V100, H100 MIG), peaking at 127 concurrent GPUs; templated per-scene jobs and packaged baselines as Apptainer images.
    • Agentic HPC ops: built a Claude Code setup (48 skills, 6 subagents, safety hooks) that submits, monitors and fixes Slurm jobs over SSH; parallel sessions in separate worktrees opened 281 PRs, 227 merged.
    • Throughput: cut CogNVS evaluation from about 3.5 h to about 21 min per scene by batching across V100s.
    • Methods: MapAnything-GS (MapAnything SfM initialization → 3D Gaussian Splatting); as an RA, dynamic-scene 3DGS (optical-flow static/dynamic split, video-diffusion refiner) and PointNeRF++ experiments.

Research

Error map of a Flash3D render aligned with test-time optimization only
Error map of the same Flash3D render aligned with PnP plus test-time optimization

WACV 2027 · First author

Aligning Misaligned Benchmarks for Single-View Novel View Synthesis

Yang-Che Tseng, Erqun Dong, Jeong-gi Kwak, Gopal Sharma, Shakiba Kheradmand, Stephan Garbin, Hossam Isack, Kwang Moo Yi

Conventional protocols compare rendered views that do not match the ground truth. We align each reconstruction to the ground-truth frame with image matching and PnP, then evaluate only co-visible pixels. Drag to compare error maps of the same Flash3D render.

  • Novel view synthesis
  • PnP-RANSAC
  • Benchmarking

Contact

Email is the best way to reach me. I am looking for ML infrastructure, LLM infrastructure and forward deployed engineering roles.

jeff@yangchetseng.com

Taipei, Taiwan