Xiyao Wang

ECE MEng student at the University of Waterloo, focused on ML systems performance — GPU profiling, inference, and the hardware underneath.

Open to ML systems co-op roles from May 2027

Selected work

GuanZero

2026

Self-play reinforcement learning for Guandan, a four-player card game, trained from random weights with no human data. Profiled the PPO rollout and added CUDA Graphs, batched attention, and a Triton KV-cache kernel for 3.48× rollout throughput on an RTX 4090 with bitwise-identical results. C++20 rules engine runs 190k decisions/s per core, verified against a Python oracle over 10M fuzz rounds.

C++20 · PyTorch · CUDA · Triton · pybind11

Runs in your browser Four seats playing a round on a JavaScript port of the rules, with simple scripted bots standing in for the trained Transformer.

Agent Call

2026

Open-source MCP server that lets an AI assistant make phone calls on your behalf — plan the call, confirm before dialing, follow along live, and get the result back. Built on OpenAI Realtime and Twilio. ~300 stars on GitHub.

Python · MCP · OpenAI Realtime · Twilio

Simulation A scripted replay of the plan → confirm → call → result flow.

Indoor UWB Localization

Capstone · 2026

Indoor positioning for a sub-250 g quadrotor. Interrupt-driven nRF52833/DW3000 firmware on Zephyr streams timing data to a Python TDoA engine with an EKF, visualized live in a React UI. Cut packet dropout from over 20% to about 3% and reached 0.455 ns synchronization RMS.

Zephyr · C · Python · React

Runs in your browser The TDoA solver and constant-velocity EKF, ported to JavaScript, tracking a simulated 3D flight through the 2 × 2 × 1 m test cell. Drag to rotate.

More projects

16-bit multi-cycle CPU

ECE243 · 2023

Verilog processor on a DE1-SoC FPGA. Single-bus datapath with a six-step control FSM, ALU, and Z/N/C flags driving conditional branches, with memory-mapped LEDs and switches.

Verilog · Quartus · ModelSim · DE1-SoC

Runs in your browser A cycle-accurate model of the control FSM in proc.v, running the original lab programs. Pick the LED counter and flip the switches.

Transformer NMT

French→English translation Transformer; 25.5 test BLEU-4.

IllustrationGreedy decoding with cross-attention.

λ-weighted DPO

Listwise, multi-objective extension of Direct Preference Optimization.

Live mathMove λ to reweight the objectives.

Canvas MCP

MCP server giving agents read-only access to Canvas LMS courses and assignments.

SimulationAn agent's tool calls, replayed.

2D Subway Surf

A Subway Surfers–style game written for the ARM Cortex-A9.

PlayableA JavaScript remake. Arrow keys or swipe.

Experience

Johnson Controls

May 2024 – Aug 2025

Junior Hardware Designer · 16-month co-op · Toronto

Hardware/software integration testing for the IQ Panel 5 across firmware, PowerG RF, and test software. Validated bidirectional RF communication to roughly 1 km range, and trained an edge glass-break detection model reaching a 98% true-positive rate.

Education

University of Waterloo

2026 – 2028

MEng, Electrical and Computer Engineering (Co-op) · Computer Architecture

University of Toronto

2021 – 2026

BASc, Electrical Engineering · Minor in AI Engineering · Dean's Honour List

Music

I produce electronic music and DJ in my spare time.