Brian Won

Brian Won

About

Hey! I'm Injong; cs grad from University of Toronto. I'm interested in ml-inference and LLM optimization

My focus spans low level serving performance and distributing training and vLLM benchmarking to memory bandwidth efficiency and cluster scale compute utilization.

Technical Skills

Languages

  • Python (OOP, multithreading)
  • C / C++
  • Java
  • JavaScript
  • SQL
  • HTML / CSS

Frameworks & Tools

  • React, Next.js
  • Express, Node.js
  • Git, Docker
  • REST APIs
  • Linux, XML

Specialties

  • Deep Learning Inference
  • Latency Optimization
  • Distributed Systems
  • Linear Programming
  • Network Security

Featured Projects

Adaptive LLM Inference Engine

600+ tok/s · P99 <3s · 32 concurrent
  • Production-style OpenAI-compatible inference gateway built on vLLM with continuous batching and PagedAttention KV-cache management.
  • Multi-GPU tensor-parallel inference with NCCL; benchmarked INT4 quantization and speculative decoding under production workloads.
  • Prometheus + Grafana dashboards tracking TTFT, KV-cache pressure, GPU utilization, and P95/P99 latency in real time.
  • Reduced inter-token latency by 22% and model-weight memory by 65% through systematic profiling and optimization.
vLLM CUDA Python Prometheus Docker

Needle DL Engine

From-scratch · C++ CPU + CUDA GPU
  • Deep learning framework built from scratch — reverse-mode automatic differentiation engine, computational graph, and N-dimensional array runtime.
  • Custom C++ CPU backend and CUDA GPU backend with tiled matrix-multiplication kernels and memory-efficient tensor operations.
  • Trained CNN, RNN, LSTM, and Transformer (multi-head attention, prenorm residual) models end-to-end on MNIST, CIFAR-10, and Penn Treebank.
C++ CUDA Python PyTorch

MiniLLM — Pretraining & Kernel Work

Flash Attention 2 · DDP · FSDP
  • Implemented Flash Attention 2 kernel in Triton with activation checkpointing to reduce memory I/O under large-context workloads.
  • Distributed training with NCCL-backed Distributed Data Parallel (DDP) and Fully Sharded Data Parallel (FSDP) across multi-GPU systems.
  • Built GPU profiling and benchmarking harness measuring kernel throughput, GPU utilization, and memory bandwidth.
  • Ran scaling-law experiments to characterize compute-optimal model sizing.
PyTorch Triton CUDA NCCL Python

EdgeRAG

Interactive Demo Available
  • Edge-optimized RAG system for financial analysis, focusing on retrieval accuracy, latency, and response grounding.
  • Iterating on chunking heuristics and metadata-aware retrieval for financial text.
  • Comparing standalone LLM outputs vs retrieval-augmented responses for factual consistency.
  • Exploring memory usage, model size trade-offs, and inference performance at the edge.
Python Vector DB RAG FastAPI React

Performance Bid-Ask Order Book

Interactive Demo Available
  • Sub-500ns price-time priority matching engine in C++17 with lock-free order ingestion and cache-line aligned structs.
  • Profiled via Google Benchmark and perf to eliminate L1 cache bottlenecks on the critical match path.
  • Dynamic order book with fixed-size array + linked list for efficient maintenance and retrieval.
C++ Prop Trading L2 Order Book WebSockets

Experience

2026

Linkedin

training model

2025

University of Toronto

Architected end-to-end ML pipeline observability using Prometheus and Grafana for platform serving 100k+ concurrent users

2023

RBC Capital Markets

QTS team:Developed latency optimization tools for trading infrastructure using stochastic modeling and data-driven analytics.

2019–2020

IBM Watson

Enhanced integration testing platform for the DB2 team, improving reliability and coverage of automated test pipelines.

Teaching

Teaching Assistant
University of Toronto
2023–2025
  • CSC373 — Algorithm Design, Analysis & Complexity
  • CSC384 — Introduction to Artificial Intelligence
  • CSC2209 — Network Topology (Graduate)

Education

Bachelor of Science in Computer Science

University of Toronto · Class of 2025

Specialization: ML & Systems