AI Product Manager · 18 years in product
I've spent 18 years building products across edtech, fintech, document automation and analytics platforms, launching several startup products from zero and leading AI teams along the way. Outside work, whenever I spot a problem, I build an AI app around it to learn what holds up. This site is where those experiments and notes live.
Most of it at B2B startups, building products from zero.
Helping a US-based SaaS company launch key AI features in a rewrite of its product, opening up a brand-new customer segment.
Led a 15+ person team building Hobasa.io, an AI platform for compliance, document intelligence and finance analytics, on multi-agent and RAG workflows.
93% faster document processing · 60% shorter compliance reviewsTook CheckanInvoice from idea to revenue and pivoted it to ML fraud detection. Owned Renepay.sg, a B2B card-payments platform built with Citi and Visa.
$6-7M weekly volume · acquired by CardUpBuilt DocMode.org and DepIndia.com, e-learning and certification platforms for thousands of doctors and dentists in India.
When I spot a problem, I build an AI app around it and take it far enough to see whether it makes sense. Superfit is live on Google Play; the rest are demos and code you can try.
Log a meal by speaking in any language, Hindi and Hinglish included, and get macros instantly. Targets adjust to your sleep and workouts through Health Connect. Everything stays encrypted on the phone.
What I learned: multilingual voice input removes the biggest drop-off point in nutrition apps, which is search.
Pulls structure and answers out of messy documents. I use it to test new document-AI ideas.
A voice agent you practise speaking with.
A private workspace where agents do the legwork of market research.
Turns receipts and statements into categorised expenses.
Drop in data, get a dashboard you can explore.
Building is cheap now. Knowing exactly which problem you're solving, and for whom, is the hard part.
AI features fail in ways normal tests miss. I decide how we'll measure "good" before the first prompt is written.
Many business tasks are better served by structured outputs, or by models that aren't LLMs at all.
I started as a business analyst in 2008 and have spent most of my career since taking products from an idea to something customers pay for, mostly at startups in fintech, SaaS and EdTech.
For the last five or six years I've been focused on AI, and for the last two or three on generative and agentic systems. I build hands-on because it keeps me honest about what models can do, where latency and cost decide the product, and what evals catch.
Outside product, I've published a Hindi poetry collection, Kuch Hakikat Kuch Falsafe, which became an Amazon bestseller.
If you're working on similar problems in AI products, I'd like to hear from you.