Case studies from 6 years of building: foundation models and evals at Microsoft Research, consumer AI on Windows, rider flows and map infrastructure at Uber, healthcare software, a startup, a hackathon win, and a lot of teaching.
I define what Quine's model should be able to do, lead the evals that measure it across mid-training and post-training, and turn the results into the roadmap, the data strategy and the compute plan.
Copilot was bad at fixing PCs because it couldn't see the PC. I owned diagnostics in Copilot Tasks, which lets Copilot write and run code locally, with the user's consent, so it troubleshoots from real system state instead of guesses.
Try this registry edit. A forum thread says it works on most versions.
forum · 2019Set it as a device policy so it covers every managed laptop.1 Pilot it on one group first.2
1 · official docs2 · your IT guidanceEnterprise customers were getting help answers no better than a web search or a generic chatbot. I connected Copilot to trusted sources and built the LLM-as-judge evals that decided when an answer was good enough to ship, raising session success 7 points.
Recall is the on-device memory for Copilot+ PCs. I built its onboarding flows and its retail demo, ran guided search experiments to help people refine their searches, and tested ideas in my own user studies with interfaces I prototyped myself.
On UberX Share I built rider flows that explain shared rides before people cancel over them, brought server-driven UI to our microservice and pitched it to the wider shared rides org, and cut MapStore's average query time 40%.
At PointClickCare, which makes healthcare software for senior care, I worked on the provider portal, built its shared UI component library, and made the CI pipelines 30% faster.
Scout was a unified search engine for personal data, with a home screen that pulls your calendar, focus timer and news into one place. We took it from idea to MVP and interviewed with Y Combinator, Neo, Pear VC and Antler.
Dots fights loneliness by turning your friends' free time into plans. It surfaces opportunities to see the people in your circle and helps you pick something to do together. Hackathon winner, UXL 2024.
from earsketch import *
setTempo(120)
fitMedia(drums, 1, 1, 9)
fitMedia(bass, 2, 3, 9)
for bar in range(1, 9, 2):
makeBeat(clap, 3, bar, "0---0---0-0-0---")With BridgeTO and Amazon I create the curriculum and run hackathons where students from underserved Toronto communities learn to code by remixing songs in Python.