# Henry Vu

> Machine Learning engineer, AI researcher, and Computer Science graduate student
> at the University of Texas at Dallas (UT Dallas / UTD) specializing in ML Systems (MLSys).
> Founding ML engineer at eXRealityAI. Research focus: efficient LLM serving and training,
> computer vision, deep learning, reinforcement learning, and AI engineering.
> Based in Dallas, Texas. Newsletter with 50,000+ readers on AI research and engineering.

> For full detailed content, see: https://henryvu.io/llms-full.txt

## Links

- Portfolio: https://henryvu.io/
- Blog: https://www.henryvu.blog/
- GitHub: https://github.com/HenryVu27
- LinkedIn: https://www.linkedin.com/in/henry-vu27/
- X/Twitter: https://x.com/HenryVu27

## About

Henry Vu is a Machine Learning engineer and AI researcher. He is a Computer Science
graduate student at the University of Texas at Dallas, focusing on Machine Learning
Systems (MLSys), making LLM/GenAI serving and training more efficient. He graduated
summa cum laude from the University of Alberta, where he was awarded the Alberta
Graduate Excellence Scholarship (AGES).

Henry is a founding ML engineer at eXRealityAI, where he builds production AI systems
for mixed reality including agentic voice systems, multimodal XR applications, and
the published Meta Quest app GardenXR. He is also a computer vision engineer at
ThorMed Innovation, working on NIH-funded medical imaging research using self-supervised
learning and model quantization for edge deployment.

Previously, Henry conducted research at Amii (Alberta Machine Intelligence Institute)
on online algorithms and multi-armed bandits, developing learning-augmented algorithms
that are both robust to worst-case inputs and consistent when given good predictions.

## Published Apps

- [GardenXR](https://www.meta.com/experiences/gardenxr/24200709416226235/): AI plant care assistant in Mixed Reality for Meta Quest. Point, scan, and get instant plant identification, health analysis, and custom care tips. Built using computer vision, Gemini 2.5, YOLOv9, and mixed reality. Free on Meta Quest Store. Rated 5/5 stars.

## Current Roles

- **Founding ML Engineer at eXRealityAI** (Aug 2025): Built voice-to-voice RAG system (Whisper STT, hybrid BM25 + FAISS retrieval, Mistral-7B-q4, Kokoro TTS) achieving sub-10s latency with local inference on NVIDIA Jetson Orin AGX. Engineered NLP pipeline with EmbeddingGemma-300m, BAAI cross-encoder reranking, and intent-aware retrieval reducing processing by 40-70%.
- **Computer Vision Engineer at ThorMed Innovation** (Feb 2025-Present): Led domain-aligned transfer learning for NIH-funded bladder segmentation. Pretrained U-Net and SimSiam SSL encoders on 9,200 ultrasound images, achieving 95.99% Dice. Enabled edge deployment via 4-bit PTQ with 7x robustness improvement.
- **Teaching Assistant at UT Dallas** (Jan 2025-Present): Algorithms and Data Structures, mentoring 100+ students.

## Blog Posts (at henryvu.blog)

- [What Fills the Context Window: A Guide to Context Engineering](https://www.henryvu.blog/series/ai-engineering/part1.html): Comprehensive guide to context engineering for LLMs. Covers the seven components of a context window, four strategies (write, select, compress, isolate), failure modes, and token budgets. 25 min read.
- [Multi-Armed Bandits: Foundations](https://www.henryvu.blog/series/bandits/part1.html): Deep dive into the exploration-exploitation tradeoff. Covers regret, stochastic bandits, concentration inequalities, Explore-then-Commit, and UCB algorithms. Includes interactive visualizations.

## Research & Papers

- [Domain-Aligned Transfer Learning for Ultrasound Bladder Segmentation](https://github.com/HenryVu27/ThorMed/blob/main/Thormed_ISBI_New.pdf): NIH-funded research on SimSiam self-supervised pretraining, U-Net segmentation, 4-bit quantization for edge medical devices.
- [Modeling Political Sarcasm](https://github.com/HenryVu27/PoliticalSarcasm/blob/main/report.pdf): DistilRoBERTa fine-tuning for sarcasm detection in political text.
- [EEG Decoding: A Multi-Modal Approach](https://github.com/HenryVu27/EEG-Multimodal-Decoding/blob/main/DecodingEEGusingTransformer.pdf): Brain signal classification fusing convolution + self-attention with LLM embeddings.
- [A Survey of Geometric Set Cover](https://github.com/HenryVu27/Geometric-Set-Cover/blob/main/HenryVu_Project_CG%20(1).pdf): Approximation algorithms and PTAS approaches.

## Notable Projects

- **Agentic Framework for Suspect Detection**: Multi-agent AI pipeline using LangGraph and Gemini for automated suspect identification with FAISS vector search.
- **Voice-to-Voice RAG for Mixed Reality**: Production RAG system with sub-10s end-to-end latency on NVIDIA Jetson, combining speech recognition, hybrid retrieval, LLM generation, and speech synthesis.
- **Online Algorithms with Machine-Learned Predictions**: Research at Amii on learning-augmented algorithms for online conversion, knapsack, and bipartite matching.
- **Multi-Armed Bandits for News Recommendation**: Applied UCB, Exp3, Thompson Sampling, and Gittins Index to Yahoo's news article recommendation dataset.
- **ViT and Contrastive Representation Learning**: Vision Transformer with contrastive learning for image classification.
- **VAE and Diffusion for Image Generation**: Variational autoencoders and diffusion models for generative deep learning.

## Expertise

- ML Systems (MLSys): Efficient LLM/GenAI serving and training, model quantization, edge deployment (NVIDIA Jetson), inference optimization
- Deep Learning: PyTorch, Vision Transformers, U-Net, self-supervised learning, contrastive learning, VAEs, diffusion models
- Computer Vision: Medical image segmentation, object detection (YOLOv9), plant identification, OpenCV
- Reinforcement Learning: Multi-armed bandits (UCB, Exp3, Thompson Sampling, Gittins Index), online learning, exploration-exploitation
- AI Engineering: RAG systems (BM25, FAISS, hybrid retrieval), context engineering, agentic workflows (LangGraph), LLM APIs, voice AI
- NLP: Transformer fine-tuning, embedding models, cross-encoder reranking, sarcasm detection
- Online Algorithms: Online optimization, primal-dual framework, learning-augmented algorithms, competitive analysis

## Education

- M.S. Computer Science, University of Texas at Dallas (2024-Present), focus on MLSys
- B.Sc. Computing Science, University of Alberta (Summa Cum Laude, AGES Scholar)

## Awards

- Alberta Graduate Excellence Scholarship (AGES)
- Summa Cum Laude, University of Alberta

## Research Affiliations

- Alberta Machine Intelligence Institute (Amii): Online learning and algorithms research
- SODALab, University of Alberta: Multi-armed bandits and reinforcement learning research
- ThorMed Innovation: NIH-funded medical imaging research

## Frequently Asked Questions

Q: Who is Henry Vu?
A: Henry Vu is a Machine Learning engineer and AI researcher, currently a CS graduate student at UT Dallas specializing in MLSys. He is also a founding ML engineer at eXRealityAI and a computer vision engineer at ThorMed Innovation.

Q: What does Henry Vu research?
A: Henry Vu's research spans ML Systems (efficient LLM serving/training), computer vision (medical image segmentation), reinforcement learning (multi-armed bandits), and AI engineering (RAG systems, context engineering, agentic workflows).

Q: Where did Henry Vu study?
A: Henry Vu graduated summa cum laude from the University of Alberta with a B.Sc. in Computing Science, and is currently pursuing an M.S. in Computer Science at the University of Texas at Dallas.

Q: What has Henry Vu written?
A: Henry Vu has authored research papers on ultrasound bladder segmentation (NIH-funded), political sarcasm detection, EEG decoding, and geometric set cover. He also writes technical deep-dives on AI engineering and multi-armed bandits at henryvu.blog.

Q: How can I contact Henry Vu?
A: Through his portfolio at henryvu.io, on X/Twitter at @HenryVu27, on LinkedIn at linkedin.com/in/henry-vu27, or on GitHub at github.com/HenryVu27.