Projects

AutoGrader — Automated Math Assignment Grading System

Machine Learning Engineer | CSC490 Capstone Design Project, University of Toronto | January – April 2026

• Collaborated with a team of 4 to build an end-to-end pipeline that automates grading of handwritten math assignments: YOLOv8 for strike-through detection, Gemini Vision for PDF-to-LaTeX OCR, and AWS Bedrock (Nova Pro) for rubric-based LLM grading.
• Architected the entire cloud infrastructure on AWS (ECS Fargate, Lambda, RDS, S3, ECR) with Terraform IaC, Dockerized services, VPC networking, and CI/CD for dev and production environments.
• Owned the LLM grading component: designed the Nova Pro pipeline and implemented a meta-LLM prompt optimizer with LLM-as-a-judge to iteratively refine grading policies and minimize MAE against human graders.
• Built confidence scoring that routes each submission to auto-grade or human review — deciding when the model should not be trusted to act alone.

RiboFold — Ribozyme RNA 3D Folding Prediction

Team Lead | University of Toronto Machine Intelligence Student Team | October 2024 – March 2025

• Led a team of 6 developers to research and develop a model that predicts 3D RNA folding of ribozymes.
• Won the Cross-Disciplinary Impact Award at the Canadian Undergraduate Conference on AI 2025 amongst 70+ projects.

Eco-Score Calculator (Partnership with Cash App)

Developer | University of Toronto, Focus in Technology Leadership Program | September 2024 – January 2025

• Developed an open-source web platform that calculates environmental impact scores for Cash App credit card transactions based on merchant ESG ratings and purchase amounts.
• Collaborating with Cash App engineers to integrate sustainability metrics into personalized financial insights.

Aercoustics — Construction Audio Classification

Researcher | University of Toronto Machine Intelligence Student Team | February – August 2024

• Fine-tuned pre-trained audio transformers with audio data from Aercoustics Engineering Ltd and improved multi-label audio classification from 62% to 85%, surpassing models by external companies.
• Performed physics-informed audio data augmentation and worked with clients directly to tailor the pipeline towards client needs.