Henry Vu

Henry Vu

Machine learning engineer

Hey, I'm Henry. I'm a machine learning engineer at eXRealityAI in Dallas, and I'm finishing my M.Sc. in Computer Science at UT Dallas, expected Summer 2027. Before that I did my B.Sc. Honors in Computing Science at the University of Alberta.

At eXReality I build XR Auto Service and Smart DVI, a spatial AI platform for auto shops. It puts an agent on a technician's headset that guides them through a repair or an inspection and answers questions from a repair knowledge graph. I lead engineering on Smart DVI and own its AI layer, and I shipped GardenXR to the Meta Horizon Store.

On the research side I work with Dr. Katherine Brown at UT Dallas on NIH-funded ultrasound segmentation, and earlier I did online algorithms and bandits at Amii. I write about all of this at henryvu.blog.

Papers

  1. Domain-Aligned Pretraining: Addressing Data Scarcity in Ultrasound Bladder Segmentation Henry Vu, Andrew Vu, Eric Chou, Gregory R. Thoreson, Katherine G. Brown. 2025. PDF Code Write-up
  2. Modeling Political Sarcasm with DistilRoBERTa Henry Vu. 2025. Report. PDF Code
  3. EEG Decoding: A Multi-Modal Approach Using Transformers and LLM Embeddings Henry Vu. 2023. PDF Code

Where I've worked

  1. Machine Learning Engineer, eXRealityAI

    Aug 2025 to now
    • Lead engineer on Smart DVI, a Meta Quest 3 inspection app in production at auto-repair shops. I direct its development and own its AI layer, from photo grading to the agents that run on the headset, VIN reading and voice.
    • Built the evaluation behind its photo grading, measuring models on real inspections and using what that showed to decide what ships.
    • Built XR Auto Service end to end, a Meta Quest 3 app that guides a technician through a repair, from the headset app to the knowledge graph and retrieval behind its assistant.
    • Shipped GardenXR to the Meta Horizon Store and built an offline voice-to-voice assistant.
  2. Computer Vision Engineer, ThorMed Innovation

    Feb 2025 to now
    • Modeling lead on an NIH-funded project to segment the bladder in clinical ultrasound, so patients can be monitored without a catheter.
    • Pretrained U-Net encoders on 9,204 thyroid and breast ultrasound images, supervised and with SimSiam, before fine-tuning on bladder images, reaching 96 percent Dice.
    • Quantized the models to 4-bit weights for portable ultrasound devices. Domain-pretrained encoders kept 75 to 93 percent Dice, while ImageNet-initialized ones fell to 11 to 71 percent.
    • Built the pipeline that extracts and cleans frames from clinical ultrasound video.
  3. Teaching Assistant, UT Dallas

    Jan 2025 to May 2025
    • Algorithms and data structures: weekly office hours, seminar and exam material, and technical reviews for more than a hundred students.
  4. Undergraduate Research Assistant, SODALab

    Dec 2023 to May 2024
    • A theoretical and empirical survey of multi-armed bandits across stochastic, adversarial, Markovian, and restless settings, with Dr. Xiaoqi Tan.
    • Simulations comparing UCB, Exp3, Thompson sampling, and the Gittins index, written up here.
  5. Teaching Assistant, University of Alberta

    Sep 2023 to May 2024
    • Algorithms and data structures for a few hundred students: office hours, weekly problem-solving sessions, and grading.
  6. Undergraduate Researcher, Online Learning, Amii

    Apr 2022 to May 2023
    • Online optimization with the primal-dual framework: conversion, knapsack, and bipartite matching problems where decisions are made before all the input arrives.
    • Algorithms that use machine-learned predictions to beat worst-case bounds when the predictions are good, without losing the guarantees when they are not.

Projects

  1. Smart DVI2025
  2. XR Auto Service2025
  3. Ultrasound Bladder Segmentation2025
  4. DatDai2026
  5. ADHD Coaching Agent2026
  6. GardenXR2025
  7. Bandits and Online Algorithms2024
  8. ML Interview Practice2026
  9. Agentic Framework for Suspect Detection2026

Contact

I read every message. Drop me a line and I'll get back to you soon.