Evaluating agentic AI at scale

Applied Scientist

Microsoft · GitHub (Vancouver, BC)

• Designed a multi-axis taxonomy and evaluated LLM classifiers for Copilot agent interactions, reaching 90%+ consistency on each intent axis and enabling offline-vs-online evaluation gap analysis.
• Adapted intent classification to multi-turn sessions using silver labels from a stronger-LLM panel calibrated with human feedback, then trained a GNN-based model that preserved accuracy at roughly 2% of the original classifier's cost and latency.
• Analyzed ~100k Copilot sessions with the Office of the CTO, identifying distribution gaps between offline evaluations and online usage that informed benchmark-instance selection for agent hillclimbing.
• Designed and built an outcome grader for task completion and user dissatisfaction, adding deeper LLM-based failure analysis for high-dissatisfaction sessions.

Senior Machine Learning Scientist

Lily AI (Mountain View, CA), remote from Vancouver, BC

• Built a multimodal RAG system combining product text, images, and VLM signals, improving attribute-extraction F1 by 25% over legacy methods while cutting token usage 10% per product.
• Designed SmartQC, an automated quality-control system using 3 LLM judges and MCP-backed tools, reducing human QC cost by 30% through high-confidence attribute auto-approval.
• Fine-tuned transformer-based vision models for image-driven attribute extraction, enabling scalable visual catalog understanding.

Machine Learning Engineer

CanDry Technologies (Vancouver, BC)

• Built a ViT-based classifier for production visual quality control, achieving 85% precision and sub-30ms Android inference with TFLite.
• Developed a ControlNet synthetic-data pipeline generating 1,000 photorealistic minority-class samples, improving rare-state recall by 15%.
• Built a PyTorch Lightning and W&B experimentation platform with 500+ automated hyperparameter trials, halving experiment-iteration time.

Machine Learning Engineer

Scenebox (Vancouver, BC)

• Built and maintained Python and JavaScript workflows for LiDAR, image, video, geolocation, and time-series data using Airflow, FFmpeg, Kafka, ROS bag processors, Redis, and Elasticsearch.
• Rebalanced search and identity-management workloads by moving selected Elasticsearch-backed data into SQL storage, reducing query latency by 20%.
• Designed, implemented, and tested a SageMaker Ground Truth integration for enterprise labeling workflows, accelerating annotation turnaround by 30% for 5+ clients.
• Maintained CI/CD infrastructure and deployed Scenebox clusters into customer AWS VPCs.

Graduate Research Assistant

University of British Columbia (Vancouver, BC)

• Led the technical work of a 5-person research team developing ControlNet-style video diffusion models for controllable cardiac structure and motion in generated echocardiograms.
• Applied consistency distillation to the video-generation pipeline, reducing latency 140× from ~7 minutes to ~3 seconds per video while preserving downstream utility.
• Curated a 100k-video echocardiogram dataset and trained ejection-fraction estimation models reaching 87% accuracy in cardiac-function prediction.

Master of Applied Science, Electrical & Computer Engineering

University of British Columbia, GPA 93.1%

• Thesis: Exploring Video Diffusion Models in Echocardiogram Generation.

Bachelor of Applied Science, Electrical & Computer Engineering

University of British Columbia, GPA 89%

Academic Service

• Reviewer for IEEE Transactions on Medical Imaging (TMI).
• Reviewer for the NeurIPS 2024 Adaptive Foundation Models Workshop.