LLM Post-training
Fine-tuning and distilling LLMs (SFT, DPO, LoRA) that replaced closed models in Heidi Health's production note generation.
Melbourne, Australia
Hi, I'm
Xuetong Wu
Ph.D. in Electrical Engineering · Staff AI Engineer
LLMASRClinical AIInformation TheoryTransfer LearningMulti-modal AI
News
2026/08
Promoted to Staff AI Engineer at Heidi Health.
2026/05
Guest lecture at UniMelb (ELEN90088) on LLM post-training and agents.Slides
2025/07
Paper accepted to IEEE Trans. on Information Theory.arXiv
2025/05
Won 2023 Best EEE Ph.D. Thesis Prize at UniMelb!
2025/02
Paper accepted to NAACL 2025 (Oral, Top 15%).arXiv
2024/09
Paper accepted to JMLR (Causal Inference).arXiv
2024/07
Paper accepted to IEEE Trans. on Information Theory.arXiv
2023/08
Paper accepted to Artificial Intelligence (AIJ) (Transfer Learning).arXiv
Introduction
I am a Staff AI Engineer at Heidi Health. I build the fine-tuned large language models (LLMs) and automatic speech recognition (ASR) models behind Heidi's clinical scribe, which replaced closed frontier models in production, and I tech-lead the on-premises deployment of our clinical AI stack for hospitals that cannot use the cloud.
I received my Ph.D. in Electrical Engineering from the University of Melbourne, where I published 10+ first/second-author papers in top venues including IEEE Transactions on Information Theory, JMLR, and ACL. My doctoral research focused on information-theoretic analysis of transfer learning and statistical learning theory.
In industry, I have built production-grade ML systems for healthcare — from LLM-powered clinical decision tools to real-time ASR pipelines and scalable data platforms serving 400+ pharmacies.
I also maintain an active research collaboration with the University of Melbourne on NLP and multi-modal AI, with recent work accepted at ACL 2025 and NAACL 2025.
What I Do
Clinical ASR
Fine-tuning and distilling speech recognition for medical conversations, from English to Korean, Arabic and Urdu.
On-prem & Inference
Running the full clinical AI stack air-gapped on local NVIDIA GPUs, served with vLLM, SGLang, TensorRT and Triton.
Multi-modal & Language AI
Out-of-distribution detection across text and vision in long dialogues, and LLMs that simulate and assess L2 English learners.
Evaluation & Benchmarking
Building end-to-end evaluation frameworks, benchmarks and quality metrics for LLM and ASR systems.
ML Theory
Information-theoretic analysis of transfer and statistical learning, published in IEEE TIT, JMLR and AIJ.
Selected Publications
All publications
Can LLMs Simulate L2-English Dialogue? An Information-Theoretic Analysis of L1-Dependent Biases
Gao R.*, Wu X.*, Kuribayashi T., Ye M., Qi S., Roever C., Liu Y., Yuan Z., Lau J.
Annual Meeting of the Association for Computational Linguistics (ACL 2025), Main Conference
Accepted (CCF-A, Co-first Author) Paper
On the Generalization for Transfer Learning: An Information-Theoretic Analysis
Wu X., Manton J.H., Aickelin U., Zhu J.
IEEE Transactions on Information Theory
Accepted (CCF-A, SCI Q2) Paper
On Causality in Domain Adaptation and Semi-Supervised Learning: an Information-Theoretic Analysis
Wu X., Gong M., Manton J.H., Aickelin U., Zhu J.
Journal of Machine Learning Research
Accepted (CCF-A, SCI Q3) Paper