Akarsh Upadhyay

I build ML systems where the solution becomes the foundation for what comes next.

Right now at Microsoft, I’m working on retrieval models for ad selection. A lot of my work isn’t glamorous: data quality pipelines, optimizations, evaluation frameworks, error analysis — but getting these fundamentals right is what actually moves the needle. When the boring stuff is solid, everything else compounds. Oh, and I also train the models at scale. They work pretty well, turns out.

Before Microsoft, I shipped AI systems to millions of users at Zomato — everything from image quality scoring and personalized photo cakes to a PoC for automated ad creation. At Enterpret, I researched text similarity at scale, spending most of my time on data preparation because that’s usually where projects succeed or fail.

I hold a B.Tech in Electrical Engineering from IIT Jodhpur (2019–2023) and have a publication at ICDAR 2024 on generic document pretraining for improved document understanding.

What drives me: I care about foundational work. The kind where if you get it right, it unlocks progress for everyone who comes after — be it data, modeling, or evaluations.