About

I'm a machine learning engineer working on computer vision, multimodal models and large language models. I build systems that understand images and video, and language models that answer from evidence rather than from memory.

I care most about reliability: whether a model still works outside its benchmark, and whether a person can check what it says. That shapes how I build models and how I evaluate them.

I'm open to research collaborations in these areas. If you're working on something related, I'd be glad to hear about it.

Research interests

  • Computer vision and video understanding
  • Vision-language and multimodal models
  • Large language models and grounded generation
  • Reliable, verifiable AI

Across all of these, I'm drawn to one question: do models still behave reliably once they leave the benchmark?

News

  • 2026 Paper

    Submitted Joint Prediction of UHPC Properties Using a Feature-Tokenizer Transformer to Applied AI Letters. It's under review.

  • Jan 2025 Career

    Joined Accelx Inc. as a Machine Learning Engineer.

  • Jan 2021 Thesis

    Graduated from AUST with a thesis on arrhythmia classification using 2-D CNNs.

Selected projects

All projects →
Evaluating a Video LLM

Overcooked AAR

I had one Video LLM review real human–human Overcooked play twice, from the event log and from the gameplay video, and checked all 1,037 of its claims against the log.

Log 77% vs. video 48%, and the video claims score the same at a random time
Qwen2.5-VLClaim checkingPilot study
 Document understanding

DocuVision

I fine-tuned LayoutLMv3-large and built a learned question→answer linker that turns scanned forms into structured key–value pairs, then released the model on Hugging Face.

Pair F1 on real scanned forms from 0.23 to 0.74
LayoutLMv3AblationsFUNSD
 Speech-to-speech translation

Speech Translator

I built a translator that speaks your words in Spanish, German or Japanese in your own cloned voice, and chose every model by testing it on FLEURS.

A 100 MB CPU model beat NLLB-200 3.3B on German and Japanese
WhisperNLLB-200Voice cloning
 Local voice agent

Jarvis

I built a voice assistant that runs fully on my own hardware and uses sandboxed tools, then tested its tool calls on a held-out set written after tuning.

Held-out tool-calling accuracy from 65% to 90%
Qwen2.5-7BTool callingWhisper

Let's collaborate

I'm open to research collaborations, especially on vision, video, or language models. A short email about what you're working on is the best way to start.

Get in touch