Hey, it's Henry
About Me
My interest in math started early through local competitions and olympiad prep. What stuck wasn't the contest tricks but the process of taking messy problems, stripping them to structure, and finding where the leverage is.
For a while I leaned into theory, spending time at Amii working on online algorithms and multi-armed bandits. For online algorithms, the focus was on making them both robust to worst-case inputs and consistent when machine-learned predictions are accurate. For bandits, it was studying exploration-exploitation tradeoffs and regret bounds across stochastic, adversarial, markovian, and restless settings.
Over time I got more interested in the gap between clean theoretical properties and systems that actually run on constrained hardware. Currently I'm a founding ML engineer at eXRealityAI building agentic voice systems and multimodal XR applications.
I write occasionally at henryvu.blog.
Outside of this
Piano, cooking, catching up with DeepSeek papers, and the occasional competitive TFT session.
Where I've Worked
AR/ML Developer @eXRealityAI
Aug 2025
- Built voice-to-voice RAG system (Whisper STT → hybrid retrieval with BM25 + FAISS → 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, query rewriting, and intent-aware retrieval reducing processing by 40-70%.
- Integrated Gemini 2.5 with structured JSON schema and context-aware prompt engineering, orchestrating stateful multimodal workflows combining YOLOv9 and Wit.ai voice input across 2 XR applications.
- Mentored development teams on LLM API integration and prompt engineering for Meta Quest applications.
Computer Vision Engineer @ThorMed Innovation
Feb 2025 - Present
- Led domain-aligned transfer learning for NIH-funded bladder segmentation project. Pretrained U-Net and SimSiam SSL encoders on 9.2K thyroid/breast ultrasound images, achieving 95.99% Dice on downstream task.
- Enabled edge deployment via 4-bit PTQ with 7x robustness improvement over ImageNet initialization.
- Built automated data pipeline with PyTorch and OpenCV: extracted, preprocessed, and segmented 486 clinical images from ultrasound videos with augmentation and quality validation frameworks.
Teaching Assistant @UTDallas
Jan 2025 - Present
- Held weekly office hours, prepared seminar and exam materials, and graded coursework.
- Mentored 100+ students in Algorithms and Data Structures through technical reviews.
Undergraduate Research Assistant @SODALab
Dec 2023 - May 2024
- Reinforcement Learning: Conducted an in-depth theoretical and empirical survey about multi-armed bandit problems under settings such as stochastic, adversarial, Markovian, and restless bandits.
- Designed and implemented Python simulations to empirically validate and compare the performance of classic bandit algorithms, including UCB, Exp3, Thompson Sampling, and Gittins Index.
- Supervised by Dr. Xiaoqi Tan.
Teaching Assistant @UAlberta
Sep 2023 - May 2024
- Held weekly office hours, prepared seminar and exam materials, and graded coursework.
- Mentored 300+ students in Algorithms and Data Structures through technical reviews and weekly problem-solving sessions.
Undergraduate Researcher - Online Learning @Amii
Apr 2022 - May 2023
- Online Learning: Researched and implemented algorithms for online optimization problems such as online conversion, knapsack, and bipartite matching using the online primal-dual framework.
- Studied a new online learning framework that incorporates machine-learned predictions to develop both robust and consistent algorithms, extending beyond traditional worst-case analysis.
Projects I've Worked On
Ultrasound Bladder Segmentation
This is an NIH-funded bladder segmentation project with limited data and edge deployment constraints. I pretrained U-Net encoders on 9.2K thyroid/breast images, achieving 7x improvement in 4-bit quantization robustness for deployment on portable devices.
EEG Decoding: a Multi-Modal Approach
This project developed a novel EEG feature extraction framework for decoding noisy brain signals, combining convolution for local patterns and self-attention for global context. The extracted features were then fused with LLM word embeddings to achieve consistent improvements in classification accuracy.
Modeling Political Sarcasm
Political sarcasm in online text is incredibly subtle, relying heavily on nuanced language and context rather than tone. This project focused on building accurate detection models using feature engineering for semantic pattern recognition, and fine-tuned DistilRoBERTa to better identify political sarcasm.
TFT Rolling Odds Calculator
Above Grandmaster level, TFT becomes a game of calculated risks where small knowledge edges are increasingly valuable. I built this tool to calculate the odds of finding a specific champion based on level, pool size, and gold using Markov Chains, where each state represents the likelihood of finding X copies of that unit.