Designed and implemented human-LLM collaboration frameworks to empirically assess how conversational agents shape users' value reflection processes.
Conducted empirical studies on crowdsourced data annotation to uncover and characterize user labelling biases throughout subjective labelling tasks.
Collaborated on LLM-persona generation initiatives, developing methodologies to drive more personalized and context-aware conversational agents.
Served as a peer reviewer for various HCI and AI conferences, evaluating and providing feedback on submissions in the human-AI collaboration community.
Guided multiple cohorts of 3-4 undergraduate students through research projects in HCI and AI development.
Focused on designing tools to observe how LLM-personas can assist users with value-sensitive decision-making tasks, with a focus on value-sensitive design (VSD).
Completed formal mentor training provided by the Computing Research Association (CRA), emphasizing culturally responsive mentorship practices.
Utilized Thompson Sampling with contextual bandits to personalize flyer recommendations to millions of users, increasing click-through rate by 8.8% yielding over $176,000 in annual revenue.
Constructed ML pipelines by containerizing our models with Docker, state tracking with DVC + Hydra and model deployment with AWS Lambda.
Designed and executed region-based A/B tests using canary releases and feature-flagged microservices to validate model performance, user engagement uplift, and system reliability in real time.
Leveraged Snowflake, AWS CloudWatch and Sisense for live model performance monitoring and visualizations.
Coordinated big data analysis and feature engineering with respect to our stakeholders and clients' expectations.
Built a high-throughput 2D image-compositing pipeline in C++/OpenGL, with multi-layer alpha blending and double-buffered rendering for seamless real-time frame assembly.
Orchestrated packet scheduling using C++11 threads and IEEE 1588 (PTP) for frame-accurate Audio / Video sync.
Implemented frame-accurate, data-source-agnostic synchronization in the in-house graphics emulator, enabling dynamic rendering from live content feeds across both local and remote media.
Sustained 60 fps on Linux by optimizing draw-call batching, texture atlas packing, and memory pooling.
Developed a WebSocket-driven TypeScript/React control panel with live-reload and hot-swap capabilities, enabling clients to preview, manage, and switch multiple broadcast feeds in real time.
Leveraged Unity’s Data Oriented Technology Stack (DOTS), Entity Component System (ECS) and Burst Compiler to optimize large-scale client scenes with over 500,000 dynamic entities, boosting frame rates from 15 fps to 60 fps.
Implemented control flow algorithms for monorail and conveyor layouts in C# within Unity to optimize package transportation at runtime, enhancing customer supply chain efficiency by 37%.
Introduced state exportation and restoration to allow clients to generate warehouse snapshots dynamically and eliminating the overhead costs of generating simulated track anomalies.
Designed automated regression testing for new and existing scenes on the CI pipeline, reducing manual labour by 66%.
Interacted directly with clients to emulate distribution facilities leveraging Unity.
Conducted research with Charles Clarke in the field of neural indexing for conversational modeling in collaboration with Meta AI Research (formerly Facebook AI).
Reduced weakly supervised training time in the Standalone Neural Ranking Model by 15% leveraging the TensorFlow library.
Enhanced the mean average precision of retrieval from 28.1% to 30.2% in the dataset processing of over 2 million queries.
Extended the Apache Lucene Core to support the combination of keyword and neural indexing.
Assisted in the coordination of CS246: Object-Oriented Software Development (Bash | C++).
Delivered tutorials and addressed inquiries on Object-Oriented principles and the Unix environment, developing 200 to 1800 line programs in C++.
Increased course efficiency through the automation and optimization of back-end processes with Python and Bash scripting, reducing testing times from 1+ days to less than 1 hour.
Developed strong problem-solving skills by concurrently debugging code for 30 to 45 students on the spot effectively during office hours 4 times a week.
Generated a web application to document back-end procedures and course coordination for future employees using HTML, CSS and JavaScript.
Conducted research in the domain of self-reflection technology for understanding bias in machine teaching under the supervision of Dr. Edith Law.
Thesis focus was on developing Human-in-the-loop LLM annotation tools aimed at recognizing and mitigating labelling biases during subjective dataset labelling tasks.
Recipient of the Mathematics Domestic Graduate Student Award, granting CAD $3,000 per year towards tuition costs.