About
I am a Senior Applied Scientist at Microsoft, working on the Office Word team 🤖 My focus is on building agentic systems that are better at personalization, context management, and tool use. I also research multimodal evaluation methods and design benchmarks for agentic flows, drawing on continual learning techniques to keep improving the system.
Before Microsoft, I did my PhD at Mila and Concordia University, where I studied continual, federated, and self-supervised learning for NLP and Computer Vision 📚 under the supervision of Dr. Eugene Belilovsky.
Outside of work, I am an avid sourdough baker 🍞 If you want tips, need help debugging your bake, or are in the Redmond, WA area and would like a starter, feel free to reach out 😊
Skills
Career
- Developing evaluation metrics and framework for agentic workflows
- Implementing agentic approaches for interactive information retrieval flows
- Analyzing and optimizing agentic flows
- Developed a novel method for system prompt migration and optimization, achieving SOTA on multiple NLP benchmarks
- Implemented automatic LLM-based metric creation and metric evaluation given a target feature
- Observed product lift for prompt migration, reducing manual migration time
- Developed and helped deploy an NLP solution using BERT to automatically populate evaluation forms of long complex legal government contracts
- Implemented content highlighting to help users verify the context used by the model to reach a decision for each evaluation criterion
- Contributed to scoping new projects and writing NLP project proposals for new clients and feature requests
- Developed an EOS X-ray image processing pipeline for 3D lower limb point cloud reconstruction, including bone and prosthetic segmentation
- Implemented fine-tuned Segment Anything Model (SAM) with manually labelled data for precise segmentation
- Created a Transformer-based neural network for 3D bone reconstruction from DRR and CT scan pairs. Deployed the pipeline on local servers for practical use
- Developed a novel method based on contrastive learning and self-prediction to learn universal user embeddings
- Applied the proposed user embedding to in-production downstream tasks and observed boost in performance
- The results were published in AMLC (Amazon internal Machine Learning Conference)
- Led the ML technical support of three projects: customer service chatbot, Python code completion, and sensitive log detection
- Carried out a series of NLP tutorials on the industrial applications of large scale language models
- Conducted workshops on data ingestion, processing, and visualisation
- Implemented hierarchical span-based toxicity detection model on in-game chat data
- Collaborated on implementing data annotation procedures
- Led a group of 5 ML Scientist and Engineers on several NLP projects
- Collaborated with the product managers to plan road maps and communicate with customers/stakeholders
- Researched and implemented multi-task learning models to improve information flow between tasks and reduce resource consumption
- Researched and implemented an extractive Question Answering system for clients with minimally labeled data
- Adapted Continual Learning techniques to mitigate catastrophic forgetting and reduce training costs
- Served ML models using TorchServe on Docker images in EMR
- Quantized and pruned the ML models to minimize the resources used and maximize throughput
- Implemented parallel and distributed model training on GPU
- Implemented a dynamic taxonomy prediction for queries, and used attention mechanism to gray-box the model
- Researched and implemented Transformer-based query retrieval system for long queries
- Implemented product content embedding based on product description and visual representation
- Participate in ML recruitment events and conduct technical interviews
- Initiated and organized "paper club", a bi-weekly workshop on the recent developments in ML
- Represented Coveo at NLP events such as MILA, RALI, MTL-NLP, etc.
- Implemented parallel model building on EMR clusters
- Implemented bigram and trigram model for product recommendation: frequently bought together recommender
Education
Jan 2021 - May 2025
- PhD of Computer Science: Continual Learning (GPA: 4.0)
- Supervisors: Dr. Eugene Belilovsky
- Scholarships:
- Concordia Merit Scholarship (2021)
- Graduate Doctoral Incentive Fellowship (2021)
Sep 2017 - May 2020
- Master of Computer Science: Natural Language Processing (GPA: 3.95)
- Supervisors: Dr. Leila Kosseim and Dr. Tien Bui
- Scholarships:
- Concordia Merit Scholarship (2018 and 2019)
- Concordia University Conference and Exposition Award (2019)
- IVADO Student Grant (2020)
Sep 2013 - Dec 2016
- Bachelor of Science: Double Major in Mathematics and Computer Science
- Represented McGill University in Putnam Competition in both 2014 and 2015
Sep 2011 - May 2013
- DEC in Pure and Applied Science
- Dawson Mathematics Competition: 1st place in 2012 and 2nd place in 2011
Teaching Experience
- Comp-472: Artificial Intelligence
- Comp-339: Combinatorics
- Comp-5361: Discrete Structures and Formal Languages
- Comp-335: Introduction to Theoretical Computer Science
- Math-242: Analysis 1
- Math-255: Honours Analysis 2
- Linear Algebra
- Probability and Statistics
- Calculus 1, 2, and 3
- Linear Algebra
- Probability and Statistics
- Calculus 1, 2, and 3
- Waves and Optics
- Mechanics
- Electricity and Magnetism