I research and design how people interact with AI systems.

AI tutors are being adopted faster than anyone has measured their effect on the learner. I wanted to know whether adapting only the tutor's communication style, not its content, would change how people learn math. The obvious answer is that it should help. The honest answer needed evidence.

I built tutors that adapted their delivery to each learner: more concrete and example-driven for some, more abstract and pattern-focused for others, mapped to Felder-Silverman learner profiles and written entirely into the system prompt. That choice matters for industry: it costs almost nothing to add to a real LLM product. Forty people worked through blocks on Venn diagrams, logarithms, and parabolas while I measured quiz scores, interaction logs, subjective load, pupillometry, and EEG.

The quiz scores stayed flat. The physiology did not. Pupil dilation during tutoring was significantly lower with the personalized tutor, meaning learners processed the same material under less cognitive load. They also engaged more, skipping fewer questions and writing out more notation as they worked. A tool can leave test scores untouched and still change the experience underneath. For anyone building EdTech: if you only measure outcomes, you miss most of what happens.

ESDC was adopting AI tools faster than its people could use them well. The Workplace Mental Health team wanted AI that fit how public servants actually work, in both official languages, without creating new risk or confusion in a sensitive domain.

I tested AI systems including agents and retrieval-augmented generation, then designed personalized solutions for the team. I consulted on AI implementation strategy aligned with the psychological factors of a healthy workplace, and wrote AI training resources and reference guides so colleagues could use the tools on the network effectively rather than cautiously.

The team moved from ad-hoc experimentation to a documented, teachable practice. Reference guides meant the knowledge survived past any single person, and AI use was framed around real tasks instead of novelty.

Calgary had tested its digital services for years, but the testers showing up were not the people most affected when those services failed. Newcomers, older adults, and people with disabilities using assistive technology were largely absent from the design loop.

I led recruitment, project management, and volunteer coordination across roughly 20 researchers between the two teams, with a target of four equity groups and eight participants each. Recruiting participants with disabilities broke the original plan: none of our 600+ existing testers matched the full criteria. I put three changes in place, a referral program with incentives that opened snowball recruitment, removal of a screening rule that filtered out eligible people, and high-touch onboarding for testers who weren't answering email.

The pilot mapped 18 distinct equity challenges across seven services. Three city service teams adopted the findings (Recreation, myID, Parks), a follow-up study on people with disabilities was approved, and mobile-plus-desktop testing for the same audience became standard.

Beneva, one of Canada's largest mutual insurers, wanted to know whether real users could complete home and car quotes as smoothly as competitors advertised, and how their live flow compared on effectiveness, efficiency, and satisfaction.

I contributed to study design, moderated sessions, coded the qualitative data, and built the severity-by-impact-and-effort matrix used to rank the final recommendations. We ran twelve participants who had shopped for insurance in the past year through four task scenarios on the live product.

The home-quote edit flow broke the study: six of twelve participants failed it because the price card still read $0.00 and they never noticed the Recalculate button at the top of the page. We flagged it as the lead usability catastrophe and delivered eight prioritized fixes with severity ratings the product team could act on.