Projects

Explore a conversational AI framework that explains complex Advanced Process Control behaviour to engineers and operators in mineral processing.

Built in collaboration with Woodgrove Technologies, a mining and mineral processing company based in Toronto, this project explores a conversational AI framework for Advanced Process Control systems used in mineral processing. The framework will bring together process data, controller information, and operational knowledge to help explain complex plant behaviour and controller actions. These systems automate complex decisions across processes such as grinding, flotation, and separation, but often provide limited visibility into why certain actions are taken. The goal is to use AI to turn this complex and fragmented information into clear, evidence-based natural language explanations, helping engineers and operators better understand system behaviour and make more informed decisions.

This project explores reinforcement learning in Rocket League, training an AI agent to perform complex in-game maneuvers such as ball control, shooting, and dribbling. Using deep reinforcement learning techniques, the agent improves over time, competing against human players and demonstrating advanced gameplay strategies.

Reinforcement Learning (RL) plays a crucial role in Artificial Intelligence (AI), particularly in complex decision-making environments. Rocket League provides an ideal platform for testing and refining RL algorithms due to its dynamic, physics-based gameplay. Despite advancements in RL, awareness in both academic and professional circles remains limited. Showcasing successful RL applications in popular games like Rocket League serves as an engaging method to educate and inform the public about RL’s potential. By highlighting practical RL implementations, this project contributes to inspiring further research and development, bridging the gap between theoretical AI advancements and real-world applications.

To mitigate problems with noisy electroencephalogram (EEG) data and financially inaccessible medical-grade EEG devices, we present 2 NLP-inspired attention-based neural networks to improve classification accuracy.

About 15 million people worldwide suffer from conditions, including locked-in syndrome, that entirely restrict their movement. Currently, thought-classification models predominantly utilize professional-grade brain-computer interfaces (BCIs), costing upwards of $25,000, rendering them unaffordable for many. Furthermore, these existing models typically achieve accuracy levels ranging from 70-80%. Our primary goal, therefore, is to create a model leveraging a more accessible 8-channel EEG setup that attains accuracy comparable to existing professional systems.

Build an interpretability toolkit to investigate the internal circuits behind uncertainty, branching, premise commitment, and error recovery in reasoning models.

The team will use sparse autoencoders and causal interventions on models such as Qwen-2.5-Math and DeepSeek-R1-Distill to develop open-source tools and a diagnostic benchmark for AI safety research.

Contribute to an open-source terminal agent for AI-assisted software development.

The project offers hands-on experience with agent orchestration, MCP and skills integrations, permissions, checkpoints, testing, CI, and browser automation in a public TypeScript and Bun codebase.

The paper aims to address the challenge of unauthorized use of copyrighted music in generative music AI by proposing a framework that ensures creators receive fair compensation while maintaining transparency in data usage. Throughout our paper, we introduce methods that combine federated and split learning with privacy-preserving techniques such as digital watermarking, fingerprinting, and algorithmic similarity analysis to detect and track copyrighted material without exposing sensitive data. Their results demonstrate that integrating these detection techniques with a levy-based compensation model can significantly reduce potential litigation costs from a projected $350 million to around $22.5 million in a case study, paving the way for a more equitable ecosystem for both AI developers and music creators

This project tackles the issue of adapting copyright law for generative AI systems, and developing reliable and safe methods of detecting copyrighted music in datasets. We believe that our proposed framework could fundamentally change the way that copyright holders and AI developers work together, by introducing ethical business practices through integrating advanced watermarking and fingerprinting techniques into AI training processes. We ensure that artists receive fair compensation while safeguarding intellectual property rights. This approach not only fosters transparency in data usage but also creates a win-win scenario for both AI developers and music creators.

Develop real-time computer vision for obstacle avoidance on a flying quadcopter.

In collaboration with ARA Robotics, the project will use a camera and onboard Jetson computing to help a quadcopter perceive and navigate its surroundings without lidar.

ARES is an adaptive ensemble strategy that trains three deep reinforcement learning agents (Proximal Policy Optimization, Advantage Actor-Critic, and Twin Delayed Deep Deterministic Policy Gradient) on a diversified universe of nine ETFs spanning U.S. and international equities, fixed income, commodities, and real estate.

Reinforcement learning offers a promising framework for dynamic portfolio allocation, but individual agents tend to be brittle across changing market regimes. Over an out-of-sample test period from October 2022 to December 2025, the ensemble achieved an annualized Sharpe ratio of 1.14 and a maximum drawdown of -10.9%, roughly half that of a passive S&P 500 benchmark, with bootstrap confidence intervals confirming marginal statistical significance.

The project explored various models (3D Convolutional Neural Network (CNN), 2D CNN + LSTM, 2D CNN) to be used in an American Sign Language (ASL) classification project with video input. This model is deployed via an interactive graphical user interface so that the model acts as an educational tool for users.

This project leverages Artificial Intelligence (AI) to create an educational interface for users first learning ASL. The datasets chosen for this project were those of ethical origins and significant diversity. Existing ASL models have been shown to demonstrate bias through class imbalances and underrepresentation of minority populations. The datasets chosen and the augmentation performed through this project aimed to reduce these common biases. Ultimately, the educational model was created to best represent individuals of all backgrounds, genders, and ages. In integrating well-performing video classification models with balanced and diverse datasets, the implementation of this platform can yield significant, not-yet-achieved benefits to the Deaf community and to all ASL learners.

Built in collaboration with Project X Ltd., a Toronto-based AI and data consulting firm, MiCRA is a multi-modal content-repurposing agent that ingests long-form company content such as call transcripts, papers, and videos.

The goal is to increase the impact of existing media by making it reusable across other channels.

Built in collaboration with the City of Kingston, this project treats property valuations as strong indicators of an area's economic activity and provides a multi-model spatial econometric system to support urban planning.

Municipal urban planners often struggle to identify optimal locations for city infrastructure because of the many factors involved in evaluating economic impact. The results show that location-based factors play a significant role in residential property values, particularly in university-centered cities like Kingston, giving municipal staff a practical tool to refine planning scenarios and strengthen evidence-based policy.

NFL Predictor is a machine learning pipeline that leverages historical NFL data to predict key player and game performance props.

The goal is to uncover statistical inefficiencies in sports betting markets and validate the power of machine learning in sports forecasting.

QMIND x OpenJustice enhances the Queen's Conflict Analysis Lab's OpenJustice tool by integrating a robust Model Context Protocol (MCP) server.

It addresses the manual and often inefficient nature of legal research, which slows a lawyer's ability to build a defense from existing legal precedent. The solution is a dynamic, real-time knowledge base that an AI agent can query and leverage to surface relevant precedent quickly.

ECHO is a system that collects, parses, categorizes, and indexes a user's everyday work.

Many knowledge workers struggle to keep track of the work they completed across a given day, week, or month. The information is presented in a clean summary each time the user turns on their device, and the user can interact with their work history through an LLM.

TriAID is an image-plus-LLM triage platform that ingests DICOM studies and automatically prioritizes them for radiologist review.

It addresses a workflow gap in busy emergency and inpatient settings, where chest X-rays typically enter the worklist in first-in, first-out order, leaving severe findings to wait behind routine exams. Each study is assigned a severity label of Severe, Moderate, Mild, or Normal, ensuring urgent findings are prioritized for review.

Built in collaboration with AI4Support, a London-based company focused on transforming how enterprises access and trust their knowledge, Cognitive Chain RAG is a Retrieval-Augmented Generation system that reasons in steps rather than stopping after a single search.

It breaks down complex questions, explores connections, and delivers responses that are well researched and reliable. Together with AI4Support, the project aims to give enterprise companies AI that does not just respond but genuinely understands their knowledge base.

Ascensi is an intelligent maze-generation agent in Unity that dynamically creates solvable mazes matched to a chosen difficulty level.

The end goal is adaptive level design within Unity.

Built in collaboration with PostopCare, a Canadian AI healthcare startup, this project replaces printed post-operative handouts with digitized instructions patients access via QR code. Already proven in a clinical setting, the platform lets clinicians upload their own documents to automatically generate a new webpage and QR code.

Digitizing post-operative handouts reduces administrative workload and paper use while improving outcomes. Because current post-op instructions are largely subjective and lack a synthesized evidence base, the goal is to use AI to build a database of evidence-based handouts, each synthesized with citations to scientific literature, that supports wider clinical deployment across a range of specialties.

Beyond Bill C-27 examines the proposed Artificial Intelligence and Data Act (AIDA) and its weaknesses through legislative analysis, case studies, and comparison with international models.

AI systems increasingly affect people's daily lives, yet Canada lacks a clear and enforceable federal law governing their use. The project proposes targeted amendments to clarify definitions, strengthen accountability, and improve enforcement, aiming for a clearer and more effective framework for regulating high-impact AI systems in Canada.

DeepfakeGuard is an open-source Python library that unifies three complementary deepfake detection modalities behind a single API: a DINOv3 Vision Transformer detector, a LipFD audio-visual detector, and a training-free D3 motion detector.

AI-generated video and audio have become an everyday attack surface for fraud, harassment, and political manipulation, yet effective defensive tooling remains largely confined to research labs. DeepfakeGuard argues that open, multi-modal defensive tooling is a necessary counterweight to the democratization of generative AI.

De-biasing Robotic Actions trains standard classifiers (Logistic Regression and Balanced Random Forest) on the MLG-ULB Credit Card Fraud dataset and demonstrates baseline violations of EU AI Act fairness thresholds.

AI systems used in credit scoring and fraud detection can systematically disadvantage protected demographic groups. These findings highlight a fundamental limitation in meeting all EU AI Act thresholds simultaneously and motivate a proposed lifecycle-based bias-audit framework aligned with the Act.

This project surveys the current landscape of AI in Alzheimer's diagnosis through a narrative review while developing a prototype model using MRI image classification, with experimental results from a fine-tuned ResNet-34 architecture.

Many predictive models for Alzheimer's disease are built on a limited number of datasets, raising concerns about dataset bias, representativeness, and generalizability. Despite high accuracy, the findings underscore the ethical challenges of deploying such AI in real-world settings. To address them, the project proposes an ethical framework grounded in justice, non-maleficence, transparency, and accountability to promote equity and responsible development in AI-based diagnostic tools.

To evaluate how LLMs respond to vulnerable users, this project administered 340 synthetic prompts to ChatGPT-4, ChatGPT-5, DeepSeek, and Gemini and analyzed the outputs through thematic analysis.

The growing use of Large Language Models for non-informational purposes raises concern that their conversational style may foster emotional vulnerability. Across 1,180 responses, 80.5% contained at least one risk-associated theme, leading the project to call for stronger safeguards, clearer boundaries, and mental health-informed design.

Using modified U-Net architectures, this project builds accurate, efficient segmentation models that perform well on small, complex datasets, aiming to automate a task cardiologists and radiologists currently perform manually.

MRI image segmentation is critical for clinical cardiac workflows, but large volumes of annotated data are difficult to obtain. Cardiologists evaluate the segmentation tool to offer insight on usability and on ethical concerns such as data privacy and accountability. A qualitative analysis of specialist interviews supports an understanding of how to integrate AI-driven MRI segmentation ethically.

This project presents a general-purpose foundation model for electroencephalography (EEG) data built with self-supervised learning.

The resulting model serves as a starting point for downstream tasks, reducing the labeled data needed to build accurate, task-specific EEG models. Evaluated on a motor imagery benchmark, the approach produces statistically significant improvements over the baseline while supporting flexible electrode configurations, with implications for few-shot clinical modeling and adaptive brain-computer interfaces.

The increasing usage of artificial intelligence in MRI disease classification and diagnosis presents several ethical impli- cations related to patient privacy, data security, and responsible use. This paper will review some current use cases of AI-based MRI image classification models and propose a framework for ethics policymakers and medical information officers to ensure patient safety and responsible usage of AI in clinical settings.

This project proposes a general guideline that policymakers and clinicians are encouraged to consult before deciding to implement an AI-based software in a clinical setting. By following these guidelines, anyone looking to implement an AI-based software can ensure that it will be used ethically and will protect patient privacy and sensitive information.

Evaluate what evidence is needed before chest-pain AI models can be trusted in real emergency-department decisions.

Can an AI model look excellent on paper and still be unready for the bedside? This project asks what evidence is needed before a chest-pain model can be trusted in real emergency-department decisions. Emergency clinicians must identify acute coronary syndrome quickly while avoiding unnecessary admission, prolonged observation, and testing for lower-risk patients. AI and Machine-learning models may help, but a high area under the receiver operating characteristic curve (AUROC) only shows discrimination. It does not establish that predicted risks are accurate, thresholds are safe, results generalize, or the model fits the clinical workflow.Beyond AUROC will update the evidence on AI/ML models used for diagnosis, short-term coronary prognosis, and coronary-risk-informed disposition in adults presenting to the ED with chest pain or suspected ACS. The review will evaluate clinical readiness using current prediction-model guidance. Its findings will then shape a feasible, workflow-constrained benchmark using MIMIC-IV data, subject to access and linkage feasibility.

Develop a responsible AI evaluation system that assesses AI-generated outputs and actions before they enter real-world workflows.

The team will evaluate reliability, fairness, safety, and the need for human review while working with industry partners to shape the system around a real responsible-AI challenge.

Explore whether large language models can improve dynamic cloud resource management through context-aware autoscaling decisions.

The research will evaluate when LLM-based reasoning adds value against traditional threshold and control-policy methods, including trade-offs in performance, resource efficiency, and adaptability.

Develop a high-precision drone, sensor, and satellite-validation system for early wildfire detection.

The team will combine thermal imaging, gas sensing, onboard computing, MATLAB analysis, and satellite data to improve wildfire monitoring and prevention.

Investigate how distinct team roles influence neural synchronization during collaborative problem-solving.

The project combines experimental neuroscience and EEG machine learning to study neural coordination in realistic teamwork settings.

Build a governed AI platform that enables intelligent agents to support complex business work and informed decision-making.

Working with Electric Mind, a consulting firm in downtown Toronto, to develop a governed AI platform that enables intelligent agents to support complex business work. The project focuses on building a reliable and scalable foundation for AI, with agents designed to analyze complex business problems, coordinate information across different areas of an organization, and support informed decision-making and strategic recommendations. The team will focus on building a reliable and scalable foundation for AI, working with the Electric Mind team to integrate AI tools in high impact fields, ensuring accuracy, traceability, and security.

Study how AI-generated recruiter profiles and deepfakes influence trust and security decisions during multi-step remote job scams.

Generative AI has fueled a massive spike in remote job scams, yet current security tools fail to address how human judgment breaks down during multi-step deceptions. It remains unclear if early trust in a synthetic recruiter profile causes victims to drop their guard, or if a deepfake's visual quality matters more than the mere presence of professional branding. In collaboration with Menlo Park Intelligence, a research organization led by Fred Heiding, Research Fellow at UC Berkeley, this project uses a remote, two-stage recruitment experiment with varying AI visual fidelity. By tracking compliance and hidden metrics like hesitation, we will isolate whether early trust reduces later scrutiny, providing the foundational data needed to design better cybersecurity defenses.

Use computer vision to measure plant growth in indoor farms and help automate growth-optimization decisions.

CanopyMap addresses the lack of high-quality training data for indoor agriculture by applying existing segmentation models to camera data from an aeroponic farm.

Use monocular video and uncertainty-aware deep learning to estimate Parkinsonian motor symptom severity while communicating when the model is unsure.

The project combines pose estimation, temporal modelling, explainability, fairness, and human oversight to explore more transparent and responsible clinical decision support.

Explore whether diffusion models can render live game worlds that change theme from a prompt while players continue playing.

The project investigates real-time generative rendering conditioned on game state, prompts, and reference images instead of fixed art assets.

Build a proof-of-concept federated learning package on DCP so hospitals can train a shared machine-learning model without sharing private data.

In partnership with Distributive, a Kingston-based distributed computing development company, we are building a proof-of-concept package that lets people run federated learning on top of DCP (Distributive's compute-sharing platform). This package will enable multiple hospitals to train one shared Machine Learning model together, without ever having to share their private data with each other, or with us. Initial application focus is predicting treatment response to biologics for psoriatic arthritis. This project will give you hands-on experience researching ML models and implementing federated learning techniques. The package we produce will be made publicly available once it's completed, contributing to the open-source ML ecosystem. We'll likely be building this in JavaScript/TypeScript, so experience with either will come in handy!