Hello, I'm Prasoon
I build practical machine learning systems that survive contact with reality.
About
I work on the parts of machine learning that have to work when the paper ends: turning messy, shifting data and ambiguous goals into well-posed problems, robust pipelines, and models that actually get used.
My experience spans experimentation, causal inference, and production ML systems: defining metrics, designing experiments, building features, and partnering with engineers and stakeholders to ship data products that move the right numbers, not just the easy ones.
I care about clarity: clear problem statements, clear interfaces between models and product, and clear communication with non-technical partners. This site is intentionally minimal – a small set of projects that show how I think, not an exhaustive catalog.
Experience
-
Leading applied ML work for user-facing products: defining problem statements, designing experiments, and taking models from notebook to production in close partnership with engineering.
-
Owning metrics, experimentation, and causal analysis for core product surfaces, helping teams choose sensible metrics and avoid being misled by noisy data.
-
Working on ML and statistics-heavy projects with faculty and peers, focusing on reproducible analysis, careful evaluation, and clear write-ups.
Selected projects
A few applied ML projects from the last few years. Happy to share deeper write-ups, notebooks, or code samples on request.
Signal Ledger
End-to-end demand forecasting replacing spreadsheet heuristics
with probabilistic forecasts and uncertainty-aware planning.
Forecasting · Product impact
Signal Ledger
End-to-end demand forecasting replacing spreadsheet heuristics with probabilistic forecasts and uncertainty-aware planning.
Experiment Atlas
Unified experimentation framework standardizing metrics, power
analysis, and guardrail checks across growth and product teams.
Experimentation · Uplift
Experiment Atlas
Unified experimentation framework standardizing metrics, power analysis, and guardrail checks across growth and product teams.
Reader Studio
Content embeddings and lightweight re-ranking powering a reading
environment optimized for sustained engagement over raw click-through.
NLP · Personalization
Reader Studio
Content embeddings and lightweight re-ranking powering a reading environment optimized for sustained engagement over raw click-through.
Notes & writing
Occasional notes on applied machine learning, evaluation, and making models behave in the real world.
When your metric fights your product
A field guide to noticing when an optimization metric is quietly misaligned with the real outcome you care about.
Evaluation · Practice
When your metric fights your product
A field guide to noticing when an optimization metric is quietly misaligned with the real outcome you care about.
Designing experiments you can actually believe
Lessons from running experiments in messy products: guardrails, pre-registration, and why celebrated uplift often misleads.
Experiments · Causality
Designing experiments you can actually believe
Lessons from running experiments in messy products: guardrails, pre-registration, and why celebrated uplift often misleads.
The quiet parts of an ML system
Feature stores, data contracts, monitoring, and the unglamorous pieces that decide whether a model survives in production.
Systems · ML in production
The quiet parts of an ML system
Feature stores, data contracts, monitoring, and the unglamorous pieces that decide whether a model survives in production.
Two GitHub Accounts, One Laptop: A Clean Isolation Setup
A practical setup for isolating personal and work GitHub contexts so folder path, SSH alias, and CLI wrapper decide account automatically.
GitHub · Workflow
Two GitHub Accounts, One Laptop: A Clean Isolation Setup
A practical setup for isolating personal and work GitHub contexts so folder path, SSH alias, and CLI wrapper decide account automatically.
Contact
For roles, collaborations, or interesting ML problems, email is best. A short note about the problem space, your data, and how success is measured is especially helpful.
Elsewhere, if you prefer. I'm always happy to talk about evaluation, experiment design, or making ML systems a little more honest.