About
My PhD was on sequence models for healthcare time series — which taught me that the model is the easy part. The hard part is the data pipeline, the evaluation that doesn't lie to you, and the deployment that survives contact with reality. I've since built recommendation and forecasting systems serving millions of users, and I've learned to be deeply suspicious of any metric that only improves offline. I publish when the work is genuinely new, and I ship when it genuinely helps.
Build models that are honest offline and useful online.
Principles: Rigor · Reproducibility · Impact over novelty · Ethical ML
Experience
Applied ML across recommendation, forecasting, and healthcare — research to production
Senior ML Engineer
2021 – PresentShopify
Lead the recommendation ranking team. New two-tower model lifted conversion 7.2% and GMV measurably.
ML Scientist
2019 – 2021Layer6 AI
Built demand-forecasting systems for retail clients; reduced forecast error (sMAPE) by 18%.
Research Intern
2018Vector Institute
Sequence modelling for clinical time series.
Education
PhD, Machine Learning — University of Toronto 2019
B.Tech, Computer Science — IIT Delhi 2014
Projects & publications
Two-tower ranking at scale
Rebuilt the recommendation ranker — +7.2% conversion, 40ms p99 at 3B daily candidates. Architecture + online eval.
forecast-kit (OSS)
A probabilistic forecasting library with proper backtesting built in. 1.8k stars.
Clinical sequence models
PhD work: attention models for ICU time series. Published at NeurIPS workshop, code released.
Eval you can trust
A framework for offline/online metric alignment that catches overfit before it ships.
References
“Priya's instinct for what will actually move the metric — not just the loss curve — is unmatched. The ranking win was one of the biggest of the year.”