Vector intern builds recommender system and semantic search enhancements

OJ Onyeagwu leaned towards his computer screen, oblivious to the summer sun streaming through the windows by his desk. He ignored the view of Queen’s Park below, focusing instead on a task he’d never attempted before – hijacking WordPress.

Just months earlier, Onyeagwu was wrapping up his material engineering degree at McMaster University. Although he specialized in computational materials, a chance viewing of Robocop during the pandemic sparked his imagination and interest in AI. He started by teaching himself through videos and online resources, eventually taking classes before finally setting his sights on a future in AI. At that point “it became very clear that the Vector Institute was somewhere I had to be if I wanted to pursue a career in AI in Canada.” 

After gaining experience at a research facility building a data analytics platform, he successfully applied for a Marketing Automation Internship at Vector, eager to dive into the world of AI.

Enhancing the Partner Portal: OJ’s internship challenge

During his internship, Onyeagwu was asked to improve Vector’s Partner Portal, a content-rich website chock full of resources for Vector’s diverse sponsor community. The Portal faced two primary challenges: a generic content feed and a limited search function. That left Onyeagwu with a lot of room to improve the user experience for sponsors.  

Working closely with Vector’s AI Engineering, and Marketing and Communications teams, he developed a two-fold solution using a Flask web API and a data pipeline.

Engineering a smarter, semantic search

By building a Flask web API with Python that received requests from the Partner Portal, Onyeagwu essentially hijacked the search functionality from the WordPress website. Now, when a user searches for content, the API transforms the query into an embedding, searches for similar content in a vector database, and returns relevant results, enhanced by Coherent Re-rank, an endpoint for refining search results. This means that the portal can now provide search results that understand a user’s intent instead of just serving up keyword matches. The enhanced search function not only displays more relevant results but also displays summaries and metadata to describe each result.  

Enhancing results with a new recommender system

The new system also includes a content recommendation carousel on the dashboard, providing personalized suggestions based on user profiles. Now, when a user logs into the Portal they are greeted with content that is relevant to their AI skill level and industry. “Someone at TD will get different recommendations than something from Google or TELUS,” says Onyeagwu. This was a “game changer” for the portal’s usability. 

Finally, Onyeagwu needed a way to gather all of the new data available on the Portal and the Vector website to ensure that his Portal improvements stayed functional after his internship ended. To do this, he deployed a Qdrant vector database that stores data as arrays of numbers for efficient similarity searches. The data pipeline, running on Google Cloud Platform (GCP), gathers and updates data from sources like Vector’s WordPress API, Vimeo, and arXiv, ensuring the database remains current. This data is then converted into embeddings and added to the vector database. Thanks to Onyeagwu’s efforts, Vector sponsors will always have immediate access to the latest sponsor-exclusive and publicly available AI resources. 

Future improvements could include implementing a feedback system like Netflix’s to further personalize recommendations, prioritizing private content, and further collaborating with Vector’s AI Engineering team to enhance recommendation algorithms. Overall, Onyeagwu’s solution improves how users find and engage with content, maximizing the value of Vector’s Partner Portal for Vector sponsors.
Before finishing his internship, Onyeagwu had the satisfaction of seeing his solution implemented on the Partner Portal. He presented his project, from challenge through to technical solution, during Vector’s Demo Day.

Growing beyond the technical AI project

While OJ’s technical achievements were significant, his internship experience extended far beyond his sunny desk in Vector’s office at the Schwartz Reisman Innovation Centre. Vector’s vibrant community and numerous events offered opportunities for learning and networking that complemented his project’s success and suited his friendly, collaborative spirit.

From attending the Collision Conference to volunteering at CIFAR’s Deep Learning and Reinforcement Learning (DLRL) Summer School, and many Vector research talks in between, Onyeagwu met key AI figures and others just starting their careers. The continuous coffee chats, mentoring opportunities like the I.M.P.A.C.T. Mentorship Program, and research conversations expanded his knowledge. They also pushed him out of his comfort zone, though you could never tell as he cheered on Toronto FC with his fellow AI and soccer enthusiasts at a game during this year’s Summer School.

Now, after completing his internship, Onyeagwu is set to start a new data analytics role with CIBC, a role that perfectly aligns with his strengths in data science. 

His internship experience cemented his belief that, for anyone to build a career in AI, Vector is the place to be. “Vector is a tight-knit community, and the people here are going places. It’s worth getting to know them.”