Runhua (Henry) Zhang

Education

University of Waterloo 09/2025 – Present

Master of Data Science and Artificial Intelligence (MDSAI, Coop)

Recognized by the Vector Institute as an AI master’s program
Coursework: Optimization for Data Science, ML, AI, DL Model Deployment, SQL, CI/CD pipelines, Flask, pySpark

University of Waterloo 09/2020 – 04/2025

Bachelor of Mathematics, Honors Statistics & Honors Computational Mathematics (Distinction)

Programming

Python (NumPy, Pandas, scikit-learn, PyTorch), R, SQL (PostgreSQL), Git, Linux/bash

Machine Learning & AI

Supervised/unsupervised learning, deep learning, reinforcement learning, model evaluation, experiment design, CI/CD pipelines

Statistics & Math

Hypothesis testing, linear models, sampling, optimization

Data Systems

pySpark, Data pipelines, validation, visualization, reproducible research

Data Management Analyst 09/2024 – 12/2024

Environment and Climate Change Canada (4 mos)

Developed a machine-actionable Data Management Plan (maDMP) aligned with FAIRER standards, enabling automated validation and cross-department interoperability for government data workflows.
Automated Excel-to-JSON schema conversion in Python (pandas, NumPy), eliminating data validation errors and enabling reproducible, machine-validated data plans.
Built a React + TypeScript schema-driven webform converting JSON schemas into interactive HTML, reducing manual Excel and schema interpretation errors.

Research Assistant 05/2024 – 08/2024

Lilia Krivodonova Lab — Waterloo, ON (4 mos)

Designed and trained FCNs and U-Nets in MATLAB for CT image reconstruction, achieving 20× speedup over iterative solvers while maintaining reconstruction fidelity.
Introduced an overlapping padding strategy to reduce boundary artifacts, improving edge consistency and robustness across scan geometries.

Research Assistant 09/2023 – 12/2023

Defence Research and Development Canada — Toronto (4 mos)

Fine-tuned LLM-based generators for synthetic clinical tabular data, enabling more realistic synthetic datasets for downstream model evaluation.
Validated synthetic data using KS, total variation distance, correlation, and contingency tests, preventing misleading downstream model evaluation on unrealistic samples.
Built an R–Python evaluation pipeline for evaluation, visualization, standardizing comparison across generators and supporting reproducible model assessment.

Research Assistant 01/2023 – 04/2023

Brian Ingalls Lab — Waterloo, ON (4 mos)

Developed Python scripts for stochastic agent-based microbial colony simulations, supporting feature and parameter analysis on growth dynamics.
Performed EDA and visualization to identify modeling limitations, guiding refinements to agent rules and simulation assumptions.

Multi-Agent Reinforcement Learning for Simulated Bacterial Colonies 09/2025 – 12/2025

AI Project

Designed a custom Gymnasium multi-agent environment to investigate whether locally observing agents can produce biologically realistic colony-level behavior under physical constraints.
Implemented shared-policy PPO (actor–critic) in PyTorch for variable-size populations, introducing Dynamic Data Alignment & Credit Assignment to stabilize learning under agent birth and death.
Engineered biologically grounded observations and hybrid local/global rewards, enabling interpretable emergent division behaviors aligned with known biological thresholds.
Diagnosed and mitigated reward-shaping exploits, preventing degenerate policies and clarifying the limits of reward-based control in physical multi-agent systems.

Hull Tactical Market Prediction (S&P500) 09/2025 – 12/2025

Machine Learning Project

Conducted rigorous walk-forward evaluation on financial time-series data with a strict six-month holdout, preventing look-ahead bias and overstated performance.
Performed extensive EDA (missingness, correlation, PCA), revealing class imbalance and non-stationarity that constrain achievable predictive performance.
Benchmarked CatBoost, regime models (HMM vs. K-Means), and Temporal CNNs, showing that regime context outperforms deep temporal pattern matching in low-signal environments.
Demonstrated that models primarily learn trend effects rather than conditional return probabilities, informing model selection away from direct return prediction toward risk-aware strategies.

ASA DataFest 2024 04/2024

Data Science Project — Waterloo, ON

Analyzed a large-scale education dataset using ANOVA and EDA to identify links between textbook difficulty and student engagement, delivering award-winning visualizations and actionable recommendations.