I trained for aircraft. The maths came with me.

I did a B.Tech in Aerospace Engineering at IIT Madras and graduated in 2021 without ever going near an aircraft afterwards. What the degree actually gave me was a habit: model a system you can't fully observe, then be honest about which terms you dropped to make it tractable. That turns out to describe most of machine learning too.

The first model anyone paid me for was a StyleGAN, at a two-month stint at Zapero.AI in 2022, generating tie-and-dye patterns from a dataset I had to scrape myself before I could train anything. That order of events (no data, then data, then a model, then something that has to keep working) has repeated in every job since.

Two and a half years at IQVIA on advanced analytics: RAG systems and NLP over public health feedback, price elasticity, demand forecasting, and the CI/CD holding the pipelines together. Then Genpact, embedded in banking and capital markets analytics, where generative AI stops being a demo and starts being something with an owner, an audit trail and an on-call rotation. Then Indivia AI as founding engineer, where the whole stack was mine, and ITO Health, where it's healthcare and the stakes are different.

What I actually care about is the whole line, from data in to model trained to service shipped, and cutting a system down until it fits in one head. Most of what goes wrong in applied AI isn't the model. It's everything either side of it.

Before all that I ran publicity and branding for Saarang, IIT Madras' cultural festival: an ambassador team of fifty-odd, the first edition of the Saarang Quiz and Saarang for Schools, and a fifth more people through the gates than the year before. It was the first time I had to make something land with an audience that owed me no attention. Still the most useful thing I learned there.

Outside work: cricket, cutting my own tools down to the bone, and an ongoing argument with myself about how much software any of this actually needs. Some of that ends up on the blog.