Whether it’s to avoid buying a problem home, resolve a dispute, or get a transparent renovation quote, property inspection is a crucial step in assessing a property's technical condition.
To automate this diagnostic process, the Ontario-based company Prop.Vu partnered with Mila to create an AI system capable of directly linking photos of defects captured in the field to technical manuals.
Collaborating to Build Expertise
Formerly known as HomePorter, the company's initial ambition was to create a virtual assistant capable of supporting inspectors. Its goal was to leverage generic language models to reference inspector training materials and assist inspectors with real-time technical guidance during virtual inspections.
For Robert Caniglia, engineer and Chief Technology Officer at Prop.Vu, Mila's support was an obvious choice to strategically evolve their product: "Our motivation was to become experts in this field, and we didn't want to do it alone."
In February 2023, the startup filled out the contact form on Mila's website, which led to the hiring of their first intern to develop this technology. This 6-months hire established the core text-processing framework, laying the groundwork for all future technical developments.
Overcoming the Data Hurdle
A year later, Robert Caniglia discovered CLIP, a learning model that acts like a matching game. This algorithm looks for the perfect pair by linking an image, such as a photo of a property defect, to its textual equivalent, like a paragraph from an inspection manual.
Robert Caniglia then recontacted Mila to adapt this visual tool to his smart search platform. Prop.Vu subsequently joined the Applied Machine Learning Research Team (AMLRT) program. For six months, the company worked with Mila to structure the technological framework.
Despite partnering with home inspection leader Carson Dunlop, strict privacy requirements for customer records limited the amount of data available to train the AI. Aldo Zaimi, an Applied Research Scientist at Mila, bypassed this data shortage by using large language models to synthetically annotate over 12,000 royalty-free inspection images gathered from the web. This approach demonstrates that carefully designed data augmentation pipelines, leveraging LLMs and public datasets, can effectively enable the training of domain-adapted models.
A Valuable Seal of Credibility
After a year of development, the model delivered by the AMLRT outperformed the baseline tool by about 10%. This proof of concept provided a valuable seal of credibility for Prop.Vu. It allowed the company to secure a landmark agreement with MPAC, Ontario’s property assessment authority managing over five million properties. These AI-assisted virtual services offer a powerful opportunity to modernize the industry, increasing throughput while keeping inspectors firmly at the center of the evaluation and decision-making process.
Beyond innovation, it is the direct access to Mila's talent pool that is propelling Prop.Vu forward today. The company has hired several interns and a PhD student in a half-time capacity to continue the work initiated by the AMLRT. Building on this synergy, the partnership with Mila has been officially renewed until 2027.
For Robert Caniglia, this alliance represents the perfect bridge between two worlds: "The synergy between Toronto's financial strength and Montreal's talent pool creates the perfect ecosystem to build world-class companies."
