Heng-Ta (Xavier) Chang
Summary
Algorithm engineer building LLM evaluation, inference infrastructure and AI agents for a consumer AI product. Background in finance and computer science, with machine learning research spanning deep reinforcement learning, self-supervised learning, recommender systems and information retrieval.
Experience
Printage, Inc.
Algorithm EngineerRemote
- Human-aligned LLM evaluation — Quantified subjective quality such as emotional companionship and writing style by cross-validating rule-based detectors, persona-based LLM-as-a-Judge and human blind A/B tests; now the basis for model selection and fine-tuning sign-off.
- Trustworthy support agent — Automates triage, root-cause diagnosis, fix proposals and verification, with every fix human-approved and reversible and read/write permissions separated; ~24× faster than the previous multi-person process.
- LLM inference infrastructure — Led end-to-end production deployment of open-weight models: GPU and cloud-vendor selection, vLLM tuning for target tokens/s and RPS, cache-key redesign (~1.35× cache hit rate) and Grafana monitoring.
- User Simulator — Persona-driven simulator for automated multi-turn testing and synthetic training data.
- Multilingual localization pipeline — LLM translation that keeps character personas and terminology consistent across long-form stories.
WorldQuant BRAIN
Research ConsultantRemote
- Alpha research — Searched the alpha-expression space with optimization methods to discover quantitative trading signals.
- Hackathon instructor — Taught other research consultants at the WorldQuant Taiwan Hackathon workshop to set up automated alpha-research workflows.
FinTech AI Lab, National Yang Ming Chiao Tung University
Research AssistantHsinchu
- Efficient RL for trading (master's thesis) — Designed a hierarchical deep RL framework for pair trading, using LLM-augmented knowledge-graph embeddings and autoencoder-based surprise rewards; converged up to 5× faster (~80% less training budget) with an out-of-sample Sharpe of 2.34 on S&P 500.
- Self-supervised recommendation × E.SUN Bank — Researched self-supervised learning for recommender systems (Sep 2023 – Jan 2024); co-presented at the E.SUN year-end research showcase.
Research Center for Information Technology Innovation, Academia Sinica
Research AssistantTaipei
- CTR prediction × E.SUN Bank — Built deep-learning CTR models for a bank recommender system, tackling severe label imbalance.
Research Center for Information Technology Innovation, Academia Sinica
Summer InternTaipei
- Cross-lingual retrieval (TREC NeuCLIR 2022) — Built retrieval pipelines for Chinese, Russian and Persian (NLLB query translation, BM25+RM3 sparse, mSBERT dense and hybrid retrieval); the team placed 3rd on the Chinese track.
Education
National Yang Ming Chiao Tung University
M.S. in Finance (Data Science track)
- Thesis — Efficient Reinforcement Learning for Pair Trading via Surprise-Driven Intrinsic Rewards and LLM-Augmented Knowledge Graphs (Advisor: Prof. Szu-Hao Huang)
- Honors — Phi Tau Phi Scholastic Honor Society; Academic Excellence Award
Yuan Ze University
B.B.A. in Finance and B.S. in Computer Science and Engineering (double major)
- Double major — 205 credits in four years
- Undergraduate research (bioinformatics × deep learning) — Identified efflux proteins with a multi-window CNN over PSSM features (MCC 0.99 vs. 0.74 for the best conventional classifier); Excellence Award, departmental project competition · Write-up
- Honors — Phi Tau Phi Scholastic Honor Society; ranked 1st in department; Academic Excellence Awards
Publications
- H.-T. Chang. "Efficient Reinforcement Learning for Pair Trading via Surprise-Driven Intrinsic Rewards and LLM-Augmented Knowledge Graphs." Master's thesis, National Yang Ming Chiao Tung University, 2025.
- C.-Y. Sang, S.-H. Huang, C.-T. Chen, H.-T. Chang. "Portfolio management using online reinforcement learning with adaptive exploration and multi-task self-supervised representation." Applied Soft Computing, 172, 112846, 2025.
- Y.-Y. Ou, Q.-T. Ho, H.-T. Chang. "Recent advances in features generation for membrane protein sequences: From multiple sequence alignment to pre-trained language models." Proteomics, 23(23–24), 2200494, 2023.
- J.-H. Ju, W.-C. Chen, H.-T. Chang, C.-W. Lin, M.-F. Tsai, C.-J. Wang. "CFDA & CLIP at TREC 2022 NeuCLIR Track." Proceedings of TREC 2022.
Talks
- "Recent Advances in Reinforcement Learning for Algorithmic Trading." 17th NYCU International Finance Conference & 3rd Yushan Conference, Dec 2023.
Awards
- Special Distinction Award, AI Application Thesis Competition (Master's Division), The 18th TSC Thesis Award (Topco Thesis Award) — 2025
- Top 25%, AI CUP 2024 E.SUN AI Challenge: RAG & LLMs for Financial QA — 2024
- 3rd place (Chinese), TREC NeuCLIR 2022 — team result
Teaching
- Teaching Assistant, National Yang Ming Chiao Tung University — Programming; Artificial Intelligence for Financial Investment Decision-making; Artificial Intelligence: Theory and Practice, 2023–2024.
- Instructor, 2023 FinTech Coding Camp (NYCU) — Taught C++ and Python to 300+ high-school students and mentored fintech and AI projects.
- Teaching Assistant, Yuan Ze University — Calculus and Web Programming, 2021–2023.
Leadership & Service
- Lead Organizer, Tjalja'avus Team, YZU IC Indigenous Service Club — Recruited and trained the university team and coordinated a winter camp with two elementary schools in Taipei and Pingtung for urban–rural cultural exchange. 2022–2023
- Volunteer Mentor, Digital and Humanity Concern Association (DIHCA) — Coached teenagers to produce documentaries on local and Indigenous communities (60+ hours). 2021–2022
Certifications
- FMA+ (Advanced Financial Management Associate) — AFMA Committee, 2024
- TOEIC 830 — ETS, 2022
Skills
- Languages:
- Python, TypeScript, C++, SQL, Shell, PHP, R, MATLAB, SAS
- Machine Learning:
- Deep Reinforcement Learning, Self-Supervised Learning, Recommender Systems, Information Retrieval, Transformers, GNN, CNN, Knowledge Graphs, PyTorch, TensorFlow
- Quantitative Finance:
- Algorithmic Trading, Pair Trading, Statistical Arbitrage, Alpha Research, Portfolio Optimization, Backtesting, Financial Econometrics
- LLMs & AI Agents:
- vLLM, LLM Evaluation, LLM-as-a-Judge, AI Agents, Agent Harness, User Simulator, MCP
- AI-Native Development:
- Claude Code, Codex, Hermes Agent, Multi-Agent Coding, Closed-Loop Iteration
- Infrastructure:
- AWS, AWS CDK, Terraform, Cloudflare Workers, Docker, Redis, MySQL, CI/CD, Grafana, Langfuse
- Spoken:
- Mandarin (Native), English (Professional)