Where I've worked

TikTok — Search Engine, ML and AI Infrastructure

Re-architected the e-commerce search pipeline to integrate a generative retrieval module with LLM-based Semantic ID retrieval into the existing multi-stage cascading system, supporting real-time inverted index serving over 800M documents. Designed and owned a distributed aggregation service that decomposed retrieval and merge logic from a monolithic engine for independent scaling — enabling 2× retrieval candidate volume with 5.4% lower ranking latency and +1.4% click-through rate, serving 141M+ daily page views across 5+ regions during 60× traffic growth. Built autonomous AI agents for performance optimization (analyzing flamegraphs, writing merge requests, self-validating via metric regression detection) and cross-region latency diagnosis. Reduced search end-to-end latency by up to 40 ms P80 (5%) via an asynchronous result prepack pipeline, lifting search engagement +0.28% across the full US user base.

UC Berkeley — Cognition and Computation in Design Lab + Donald O. Pederson Center for Electronic Systems Design

Led development of a Chrome extension with TypeScript, React, and Flask integrating generative AI for visual stimuli research. Built a personalized AR rehabilitation task authoring tool using Unity on Meta Quest 3. Created a compliant EMR web system using React, Next.js, Node.js, and Firebase. Hosted cloud infrastructure on AWS, integrating CI/CD with Jenkins and Docker.

University of Toronto — Intelligent Adaptive Interventions Lab

Deployed full-stack SaaS tools using GPT-4 and LLMs to support youth in mental health and creative writing. Developed apps with React.js, NestJS, Flask, and MongoDB for iOS, Android, and Web. Led user studies and co-authored 3 peer-reviewed papers, including CHI submissions.

Huawei Technologies Canada — Distributed Data and Storage Management Lab

Implemented buffer pool with LRU eviction and page directory modules for GaussDB, Huawei's enterprise-grade distributed PostgreSQL-based database. Used consistent hashing with virtual nodes to partition pages across nodes, enabling hash-directed failover for fast replica lookup on node failure. Developed replica-assisted crash recovery so restarting nodes rebuild state from peers' in-memory buffer pools rather than disk — contributing to a 150% throughput increase on industry-standard OLTP benchmarks. Led a test group that built concurrency and failover suites validating correctness under concurrent buffer pool access, page directory updates, and multi-node crash recovery.

University of Toronto — Lee Language Lab

Conducted 100+ ML experiments optimizing translation models like mBART, mT5, and XLM-R. Boosted BLEU scores by 20% for low-resource languages through task transfer learning. Led project management across the full research lifecycle.

University of Toronto — Social Networks (Prof. Peter Marbach)

Designed and assessed six utility functions for defining social-network hierarchies, applied to over 80,000 users and 12.7M tweets. Presented findings to 100+ students and faculty at the 2022 ROP Fair.