michael zhang

uwaterloo cs '31

about

incoming university of waterloo cs student interested in systems, ai/ml, and game development. i work on machine learning research (model compression, quantization, and continual learning in llms), build games, and ship developer tools.

education

university of waterloo
bachelor of computer science (co-op)
expected 2031
marc garneau collegiate institute — tops program
ontario secondary school diploma
2026

experience

founder & owner at roblox game studio
solely handled the majority of development across 4+ published lua games (1M+ total plays): game design, gameplay and backend programming, ui/ux, debugging, performance tuning, and live updates
programmed core systems including data persistence, matchmaking, inventory management, and enemy ai, structured around a modular architecture with custom module loaders for reuse across titles
engineered server-client architecture using remoteevents and remotefunctions for secure multiplayer, with object pooling and lazy-loading to hold memory and latency low at scale
2024 - present
software developer intern at royal bank of canada
selected to build internal ai-powered tools and prototypes at rbc's toronto headquarters
summer 2026
machine learning intern at algoverse
investigated the effects of 8-bit quantization on continual learning and catastrophic forgetting in large language models, finding that lower-precision models can retain prior-task knowledge better than full precision
fine-tuned llama-3.1-8b at fp16/int8/int4 precision using lora/qlora across nlu, math, and code-generation tasks; quantized models outperformed full-precision fp16 by 8-15% on later-task accuracy, with int4 nearly doubling fp16 on code generation (40% vs 20%)
built an evaluation pipeline mapping the quantization × replay-buffer tradeoff; a minimal 0.1% replay buffer raised nlu retention after math training from 45% to 65% across all precision levels
first-authored a paper accepted to the ccfm and llm eval workshops at neurips 2025
may 2025 - dec 2025
president at computer science club
led 20+ student developers building 15+ javascript automation tools (attendance tracking, transcript scanning, club management) used by 15+ staff and clubs, saving hundreds of hours and serving 2,000+ students
designed and ran weekly algorithm workshops for 30+ members, preparing students with no prior contest experience for the canadian computing contest and other national competitions
sep 2024 - present

publications

m. zhang et al. "when less is more: 8-bit quantization improves continual learning in large language models." ccfm & llm eval workshops, neurips 2025. arxiv:2512.18934

projects

daedalus — jamhacks 10 winner (best non-genai hack)
built a browser-based 3d modeling tool driven entirely by webcam hand gestures (no headset, depth sensor, or install), using mediapipe hand tracking and three.js/webgl to map 2d hand landmarks to real-time 3d transforms such as translation, quaternion-based rotation, and scaling
implemented boolean csg modeling (union/subtract/intersect) with three-bvh-csg plus bvh-accelerated clay-style sculpting, and integrated an ai voice assistant (web speech api) for natural-language shape generation and scene styling; built in typescript/vite with a 4-person team
jamhacks 10
spinal degeneration ai classification model
built a cnn in tensorflow/keras with mst-based roi extraction to classify lumbar spine degeneration severity from rsna 2024 mri dicom scans; achieved 0.44 log loss and 92%/78% accuracy on mild/severe cases
independent research

awards

usaco — reached gold division
aime qualifier
canadian computing contest 2025, senior division — scored 54/75, above the certificate of distinction cutoff (34/75)
hosa ilc, 6th place internationally — 6th of 1,000+ teams in a health sciences team quiz bowl
deca icdc 2025 international finalist — top 20 internationally out of all deca ttdm competitors

skills

languages: python, java, c/c++, c#, javascript, typescript, lua, bash, html/css, sql

ml & research: pytorch, tensorflow/keras, hugging face transformers, peft (lora/qlora), bitsandbytes, scikit-learn, weights & biases, numpy, pandas, matplotlib

tools: git/github, linux, aws, node.js, react.js, three.js, roblox studio, vs code, jupyter/colab

contact

email: mzhang3518@gmail.com

github: github.com/festyve