Dongruixuan Li

Built an internal multimodal AI chatbot to assist investment analysts with documents search. - Developed end-to-end backend using Python, pgvector + PostgreSQL, and a vision model for image and text retrieval. - Deployed scalable infrastructure on AWS EKS using Kubernetes, ELB, and S3. - Collaborated with frontend and product to deliver features rapidly in a startup setting. - Enabled analysts to query across documents—improving research efficiency.

Worked on performance and UX improvements across backend APIs and trading dashboards. - Reduced internal dashboard load time from 19s to 3s by building a GraphQL caching layer. - Enhanced trader UI by integrating dynamic strategy controls with React + TypeScript. - Fixed animation/render bugs on internal dashboards, improving clarity during filter updates.

Worked on real-time content moderation and large-scale analytics infrastructure. - Built a middleware-style monitoring system with Go and Message Queue to validate outputs across moderation pipeline changes. - Identified and resolved data races in core frameworks, improving system reliability under 6000+ QPS. - Migrated mission-critical OLAP workloads from Elasticsearch to ClickHouse, reducing p95 query latency by 92% on 12B+ records. - Profiled and optimized high-QPS Go services with pprof, reducing CPU usage by 210 cores (27%). - Created a policy analytics dashboard using Hive, Spark, and Elasticsearch, enabling Ops to track moderation rules in real-time.

Large Language Model Research Infrastructure: - Developed a distributed high-performance web scraping platform, with ability to parallel fetch webpages via simple HTTP request or browser automation. Techniques: Go, Gin, MongoDB, RabbitMQ, ChromeDP(like Selenium), Docker Performance: The bottleneck is on all worker’s CPU (using browser automation), or on all worker’s network speed (using simple HTTP request). - Installed a Ceph storage cluster and a GPU Kubernetes cluster with mixed GPU models for lab internal use; highly fastened provisioning LLM testing instance. Pre-train a cybersecurity-related large language model based on Llama2-7B model - Reproduced a data cleaning pipeline with colleagues using Python. - Used cProfile to analyze the performance bottleneck, then rewrote the MinHash algorithm in C++ for a 100x performance increase. - Pre-trained the model using Transformers, with DeepSpeed ZeRO-2 to allow training on low VRAM, and with Flash Attention to approach a better performance and lower peak VRAM usage. Developed a vulnerable frontend of the AI Chatting app for a security camp. Techniques: Electron, React, Redux, WebSocket, Material UI

First in Class Engineering Scholarship (Term 2B) First in Class Engineering Scholarship (Term 1B) President's Scholarship of Distinction Engineering International Student Scholarship