Selected projects

Echocardiography report generation

A reimplementation of Stanford's EchoPrime that generates structured reports from multi-view ultrasound videos.

• Adapted the pipeline to UBC's proprietary clinical dataset.
• Fine-tuned a text encoder with contrastive learning on millions of echo videos and thousands of cardiac reports to improve video–report alignment.

• Multimodal learning
• Contrastive learning
• Medical imaging

LLM benchmarking framework

A provider-agnostic framework for comparing LLMs on image-based IP infringement detection.

• Integrated OpenAI, Google, Anthropic, Mistral, and AWS APIs behind one interface.
• Raised image-analysis accuracy by 5–10% with prompt optimization, multi-step reasoning chains, and grounding.
• Fine-tuned Gemini 1.5 Pro and Flash models on proprietary IP datasets, achieving a 5–10% improvement in intellectual property recognition accuracy over baseline models.

• LLM evaluation
• Prompt engineering
• Fine-tuning

Real-time image classifier

ViT and hybrid ViT-CNN models that detect images containing valuable assets.

• Collected 80,000+ images and built a Hugging Face API pipeline for training data.
• Cut latency by 75% with TorchScript compilation, knowledge distillation, and INT8 quantization.

• Computer vision
• Model compression
• Hugging Face

ROS autonomous vehicle

An end-to-end autonomous driving system in ROS, running in the Gazebo simulator.

• Deployed neural networks for perception and control.
• Trained a CNN via imitation learning to follow lanes and navigate from camera input.
• Built an OpenCV preprocessing and labeling pipeline to generate supervised training data.

• Python
• OpenCV
• TensorFlow
• ROS
• Imitation learning
• Simulation

Sleeper fantasy football assistant

A web app that connects to a Sleeper fantasy football league by username and shows standings, rosters, matchup lineups, and player stats scored under the league's own rules.

• Standings, rosters, lineups, and a searchable players tab, with projected and actual points calculated from each league's scoring settings.
• Python and FastAPI backend, Next.js and TypeScript frontend, and Postgres storage.
• CI runs strict type checking, tests, linting, and accessibility checks.
• Early development: consensus rankings, waiver and trade suggestions, and a playoff-odds simulator are planned next.

• Python
• FastAPI
• Next.js
• TypeScript
• Postgres

View source on GitHub · Live demo

This website

A static site built with Astro, with no client-side framework.

• Automatic dark and light themes, a responsive layout, and accessible markup.
• Linted in CI and deployed to GitHub Pages with GitHub Actions.

• Astro
• CSS
• GitHub Actions

View source on GitHub

Liver health assessment platform

A web-based diagnostic platform for a medical device, built in a 3-person agile team.

• C++/JavaScript backend with a WebSocket server for low-latency data exchange between device code and a browser GUI.
• Python REST API managing SQLite databases for patient and device records.

• C++
• Python
• WebSockets
• REST