Innovative Projects Realized

Exploring a Multi-modal Audio Foundational Model

Boson AI is an early-stage startup of 30 scientists. They are building large language tools for interaction and entertainment. They are seeking interns from the University of Toronto to join them in their Toronto office. The intern will work on modeling and training LLMs, understanding and interpreting model behavior and aligning models to human values. The company will provide comprehensive guidance and resource support during the internship.
During this project, the partner wants to build an AI model to understand and speak in a natural and emotionally competent manner. Humans often prefer voice communication over text. This has to do with the immediacy of the response and the amount of emotional nuance and subtext that can be conveyed beyond the actual words. Details are only obtainable through the audio (tone, laughter, pauses, volume, etc.), and humans are very good at picking this up. Computers are typically less good at this.
The project is expected to bring significant social and economic benefits by enhancing AI’s ability to understand and express emotions naturally in speech. This advancement will improve human-AI interactions, making virtual assistants, customer service bots, and social robots more engaging and intuitive.

Aston Dynamics Canada EBS project

Aston Dynamics Canada Limited (Aston Dynamics) is working on improving electro-hydraulic braking (EHB) systems for electric and heavy-duty vehicles. The company has already made significant progress in solving key issues like slow braking response, reliability, and limited braking force. They’ve developed and tested prototypes that show better braking control and efficiency, especially for electric vehicles, trailers, RVs, and agricultural vehicles. As the project continues, the focus will be on enhancing the system’s electronics, testing it in real-world conditions, and preparing it for commercial use. This project will help Aston Dynamics lead the way in making braking technology safer and more reliable, benefiting the partner organization by offering advanced, efficient, and industry-leading EHB systems.

Data Systems & ML Ops Redesign

J Squared builds and provides rugged electronic systems in several industries. FALC-AI, a division within J Squared, is focused on the vertical integration of these platforms with computer vision and AI models to enable a variety of use cases that require being
able to extract accurate data from videos and images, and in turn, use that data to enable industry specific features. As a result of this, data from edge devices is streamed into the cloud for processing. With a low number of edge devices, the performance and storage requirements are pretty minimal. But, as more devices are added, the current solution will need improvements on the data processing side to be able to operate on the data in a performant and timely manner. This project will focus on researching, testing, and implementing a variety of improvements across the full data pipeline—from infrastructure setup and query optimization to data cleaning, library evaluations, and benchmarking—to ensure robust and scalable solutions for processing data from edge devices. The primary focus is optimizing the performance of a compute-intensive stage in the pipeline, where a YOLO-based object detection model and a Person Re-Identification (RE-ID) model, OSNet, run in parallel. Since the RE-ID model is particularly resource-intensive, it’s crucial that all data processing in this step is highly efficient to maximize overall system throughput.

Fine-Tuning Large Language Models for Standards and Claim Chart Mapping

XLSCOUT Ltd. specializes in AI-driven patent intelligence, providing advanced solutions for intellectual property (IP) strategy and innovation management. The company aims to enhance the accuracy and efficiency of patent-related tasks, such as claim chart generation and standards compliance mapping, through fine-tuned large language models (LLMs).
This project addresses a critical challenge: general-purpose LLMs lack domain-specific understanding required for precise claim chart mapping and regulatory compliance. By fine-tuning LLMs for standards interpretation and automated claim chart generation, the project will significantly reduce manual effort while improving accuracy and consistency in patent validation.
The anticipated benefits for XLSCOUT include enhanced AI-driven tools that improve workflow efficiency for IP professionals, reducing time and cost associated with claim chart creation. This will strengthen XLSCOUT’s competitive position in the patent intelligence market by offering a unique AI-powered solution tailored to industry needs.
Beyond XLSCOUT, the broader IP sector will benefit from more reliable and standardized claim chart mapping, leading to improved patent litigation and enforcement. Additionally, this research will contribute to advancements in AI applications for legal and technical documentation, demonstrating the potential for LLMs in specialized regulatory domains.

La lixiviation des dépotoirs, un risque pour l’environnement?

La gestion de nos déchets est un défi de taille du 21e siècle : si la part du recyclage/réacheminement est en constante progression au Canada depuis quelques années, il n’en demeure pas moins que la vaste majorité des déchets collectés aboutissent dans des sites d’enfouissement, posant ainsi des risques de pollution pour les écosystèmes avoisinants. L’infiltration et le ruissellement des eaux de pluie sont susceptibles de mobiliser les contaminants associés aux déchets et représentent ainsi une source de pollution potentielle tant pour les nappes phréatiques que pour les cours d’eau de surface. Si la collecte et le traitement de ces eaux permettent de minimiser les impacts environnementaux, les canalisations nécessaires à une telle collecte souffrent d’encrassement limitant l’écoulement des eaux et l’efficacité globale du processus. Telle est du moins la réalité de notre partenaire ECO360 en charge du site d’enfouissement de la région du sud-est du Nouveau-Brunswick. C’est pourquoi deux objectifs communs ont été identifiés pour ce projet : 1) caractériser l’enrichissement des eaux de lixiviation en 4 éléments majeurs (Ca, Mg, Fe, Mn), 7 éléments toxiques (As, Cd, Cu, Cr, Hg, Pb et Zn) et 2 autres contaminants émergents (Ra et U) et 2) identifier la nature de l’encrassement.

Leveraging AI and IoT to predict heatwaves in Canada: A climate health initiative

This project uses AI and smart thermostats to predict heatwaves in Canada, helping protect public health. By analyzing indoor and outdoor temperature data, we aim to improve heatwave forecasting and understand how extreme heat affects indoor spaces. Using machine learning, we develop models to predict future indoor temperatures, supporting better planning and response strategies. The project benefits institutions by providing valuable data for public health officials and policymakers, helping them create early warning systems and improve heat-related safety measures, especially for vulnerable groups.

Advancing the Cellular Agriculture Industry with the Power of Raman Spectroscopy

This project will integrate an advanced analytical technique (i.e., Raman spectroscopy) and machine learning to help the partner organization monitor key cell growth parameters, streamline experimental design, accelerate culture media optimization, and ensure the consistency of media and product quality. These advanced analytical tools and workflows will allow the partner organization to improve the efficiency and precision in cell-cultured pork production, accelerating the commercialization of their products, and strengthening their position as a leader in the cell-cultured meat industry.

A Patient-Centered and Visually Attractive Toolkit Adapted to Children and Youth to Reduce Inhaler Impact on Carbon Footprint in Ambulatory and Outpatient Settings

A Patient-Centered and Visually Attractive Toolkit Adapted to Children and Youth to Reduce Inhaler Impact on Carbon Footprint in Ambulatory and Outpatient Settings
1) Introduction/Background: Carbon footprint (expressed in carbon dioxide equivalent, CO2e) is the preferred method to measure our impact on climate change. Healthcare represents 4.6% of total Canadian greenhouse gases emissions, and in this category, pharmaceuticals is making up to 25%. Pressurized Metered-dose Inhalers (pMDIs), are known to have a higher carbon footprint impact than their primary alternative, the Dry Powder Inhalers (DPIs), due to their liquified gases called hydrofluoroalkanes (HFAs). In a recent REB approved project, we completed a chart
review to examine the prescribing of inhalers in our pediatric respirology clinic, from January 1, 2022, to Dec 31st, 2022. Our research project highlighted that 92.9% of a year of inhaler prescriptions in our respirology clinic were pMDI inhalers with a high switchability potential to DPIs, if deemed clinically feasible. We estimated that switching 30% of salbutamol prescriptions from MDI to DPI could save around 16 metric tonnes of eCO2 emissions, an equivalent to 6 flights round trips between Toronto and Tokyo. As we were planning to share knowledge to our peers and influencing change on MDIs to PDIs prescription, when clinically safe, we could only find toolkits for a sustainable prescribing and usage of inhalers in the adult population but none that were adapted and patient-centered for the pediatric population. This is an opportunity to collaborate with clinicians, in a carbon stewardship model, and expand a proven carbon
reduction model into tools, adapted to the pediatric population.
2) Purpose: The purpose of this quality improvement
project is to design and share, a video, a handout and a decision tree, adapted to the pediatric population, as tools to simplify decision making processes in inhalers prescription by healthcare professionals.

PaceZero BSI 2025

PaceZero is hiring an intern to research how to integrate Artificial Intelligence (“AI”) into operations. Areas of interest will focus on, but are not limited to, leveraging technology to enhance sourcing prospective borrowers. More specifically, the intern will research the capabilities and functionality of various Ai programs such as ChatGPT, Perplexity, and tl:dv, among others. Research will explore how Ai can support various business processes: i) sourcing prospective borrowers by searching available data bases for borrowers that meet criteria, ii) integrating ai meeting notes into CRM iii) monitoring various websites for relevant news and automated push notifications iv) other minor use cases that may arise during the course of the term.

Automatically Evolving Machine Learning Codebases with Large Language Models

Machine learning (ML) is transforming industries like IT, finance, and healthcare, but the code that powers these systems is still mostly written and updated by hand. This project explores how Large Language Models (LLMs) can assist developers by predicting and suggesting code edits for ML projects.

By analyzing real-world code from public repositories, the research will develop and test an AI-driven approach to automate ML code modifications based on developer intent. The goal is to make software development faster and more efficient, helping businesses and researchers automate the process of updating codebases.

This project aims to build a benchmark dataset of ML code edits, develop an LLM-powered assistant that can suggest and apply modifications, and evaluate its effectiveness compared to human-made changes. By improving automation in software engineering, the research will help streamline ML development and reduce manual effort.

Asynchronous mesh networks for continuous methane leak monitoring

Spero Analytics is a research-driven IoT startup which builds mesh networks for automated, continuous surveillance of greenhouse gas emissions in landfills and energy facilities. As those networks are power efficient and operate using radio communication, they can be deployed in remote industrial areas with limited cellular/power infrastructure. The existing system architecture consists of a gateway connected to a number of downstream nodes arranged in a multi-hop mesh configuration and equipped with a high-accuracy methane sensor. Every hour, the nodes come online, transmit their sensor reading via their on-board antenna, then go into a “deep sleep” mode to conserve power. This existing arrangement allows us to deploy robust, scalable greenhouse gas surveillance networks that can operate for years with minimal operator intervention. The challenge, however, emerges when there is an anomalous methane reading (due to a leak) which occurs during the hour that the nodes are in deep sleep mode. In that case, the operator will not be alerted about the anomaly until the next reporting cycle, thereby introducing a delay in leak detection and repair. This project proposes the development of an algorithm which will allow the mesh network to react to those intra-duty-cycle anomalies without compromising power efficiency.

Biosignal Transformers for Advanced Blood Pressure Waveform Analysis

This project aims to develop advanced machine learning models to analyze arterial blood pressure (ABP) waveforms
from patients in intensive care units (ICUs). By using a large dataset and cutting-edge techniques like transformer
architectures and contrastive learning, the goal is to create models that can accurately predict patient outcomes, such
as ICU mortality and hospital stay lengths. This research will improve our understanding of how ABP data can be
used to support clinical decision-making and enhance patient care. The project will also benefit participating
institutions by advancing the application of machine learning in healthcare, providing valuable insights for both
academic research and real-world clinical practice.