I started at Mu Sigma in 2014, spending two years building segmentation, demand forecasting, and marketing mix models for Fortune 500 retail and airline clients. It was a strong foundation in applied analytics - learning how to frame problems, what good questions look like, and how to translate analysis into business action.

I joined Amazon in 2017 and spent nine years there. That tenure took me across payments and high-frequency commerce in India — first as a BA, then leading the analytics function — before moving to private brands in Luxembourg, and eventually the selection expansion team, where I was the sole BIE advising Director-level monthly business reviews on a ~$4B cross-border seller program across 17+ global stores.

What changed for me at Amazon was the scope. I went from building models to building systems - things that needed to run reliably, communicate clearly to many stakeholders, and hold up under business pressure. That experience fundamentally shaped how I think about analytics as infrastructure.

The current chapter is a deliberate one. After nine years inside large-scale analytics organizations, I chose to step back from full-time corporate work and build — to go hands-on with the kind of LLM-driven intelligence systems I'd spent a career scoping for others. I'm using this time to build real products and deepen my applied-AI craft, while looking for the right product or analytics leadership role to step into next. Open to consulting engagements along the way.

The most complete is Vyoman — an aerial-photography venture whose infrastructure I built end-to-end: a public site and print store (Stripe + Gelato print-on-demand), a Django ops console with an LLM intelligence dashboard, a canonical data layer that serves as the single source of truth, and a creative pipeline prototype that grades footage and drafts the story. A change in the console goes live on the site in seconds, with no deploy. Commercial launch is pending business registration.

Alongside it: Disha Analytics — restaurant intelligence via an eight-agent LLM pipeline, with one pilot report walked through with a restaurant owner in New Delhi; ReasonTrail — an AML/KYC review workflow prototype for Luxembourg with four-eyes approval, an immutable audit trail and CSSF-format export, plus a designed AI drafting layer; CircleWorks — community trust infrastructure with LLM-powered pricing over a Telegram bot, piloted in one residential community in Bangalore; Soul Spark — multi-agent life intelligence that runs entirely on-device; and Narrate — a personal audiobook system that converts any PDF or document into narrated audio with voice personality per book, adaptive chapter conversion, and lock-screen controls. Each one starts from a real problem; they range from pilots to working prototypes to tools I use daily.

Building these end-to-end has made me fluent in what actually makes AI products work: canonical data layers and intelligence dashboards, multi-agent and LLM pipelines, the discipline of keeping a deterministic core while the LLM only drafts, and local-first privacy patterns. The throughline is the same one from my Amazon years — turn messy, real-world data into structured, decision-ready output.