Performance and tracing
Finding where time goes in systems too large to watch by hand: trace optimization, performance regressions at the method level, root-cause analysis in microservices, and GPU benchmarking.
Five threads run through the group's work. Each card opens the matching papers.
Finding where time goes in systems too large to watch by hand: trace optimization, performance regressions at the method level, root-cause analysis in microservices, and GPU benchmarking.
Putting language models to work on real engineering tasks and measuring what comes out: multi-agent refactoring, log analysis on private data, generated review comments, and text-to-SQL evaluation.
Machine-learning software is still software, but it breaks and ages in its own ways. We study its exceptions, its bugs, how it gets refactored, and how to plan defect reduction.
How changes get reviewed, built and shipped: reviewer recommendation and its risks, how review feedback evolves, and whether CI acceleration can be trusted.
Where I started: how APIs change, how their users cope through migrations and workarounds, and how examples mined from public code can help both sides.
Applied project
Ethical, sovereign AI that brings climate data together in one place so that Québec municipalities can assess their vulnerabilities and build resilience to climate change. Covered in Le Devoir, May 2026.