All writing

StartupBy 4 min read

Building FinMoon AI: automate the expert process, not the answer

Why we built a personal finance decision assistant for Indian professionals, what generic AI chat gets wrong, and what getting into the Google for Startups Hub at T-Hub taught us.

Updated

On this page · 6 sections
  1. What is FinMoon AI?
  2. Most Indian professionals manage money on their own
  3. What generic AI chat gets wrong
  4. Product principles I’d reuse anywhere
  5. Google for Startups Hub at T-Hub
  6. What building it has taught me

For the last few years, my co-founder Shashank Sheela and I have watched smart, well-paid friends and colleagues make important financial decisions badly. They weren’t careless. A good decision simply takes a lot of manual work, and the work is tedious enough that people get it wrong.

That observation became FinMoon AI. This post covers the thinking behind it and the product principles I’d apply to any AI product.

What is FinMoon AI?

FinMoon AI, at finmoon.ai, is a personal finance decision assistant for busy Indian working professionals. You bring a decision, such as a home loan, a credit card, an insurance policy or where to park an emergency fund. It asks about your situation the way an advisor would, compares the options against what matters to you, and gives you a report: what to compare, which risks and hidden costs to watch, and what to do next.

Most Indian professionals manage money on their own

Good, unbiased financial advice is hard to get in India. Fee-only advisors are few for a country this size, and wealth managers mostly want clients who already have large portfolios. So most professionals handle their money entirely by themselves, and many end up under-invested, carrying expensive debt or under-insured.

When we dug into why managing your own money is hard, three problems kept coming up.

  1. Financial products are confusing by design. A home loan, credit card or insurance policy comes wrapped in jargon and hidden clauses, and the terms change every year, so one choice turns into hours of fine print.
  2. Decisions affect each other. A home loan changes how big your emergency fund should be. An insurance gap changes how much you can safely invest. Nobody maps those connections, so people optimise each decision separately and miss the interactions.
  3. The work continues after you sign. Rates move, policies get re-priced, and the product that made sense last year may not make sense today. Someone has to keep watching, and usually nobody does.

What generic AI chat gets wrong

In 2026 it’s tempting to wrap a chat interface around a large language model and call it a financial advisor. People already use general-purpose chatbots this way, asking about tax regimes, SIPs, home loans and insurance, and the answers feel right.

They’re usually incomplete, and better prompting won’t fix that. A general chatbot answers the question you asked. An expert runs a process. They ask about your income stability, obligations, dependents, liquidity needs and tax position before they say anything. Then they compare options against the criteria that matter for your situation, and tell you what to watch after you decide.

What makes an expert useful is the process behind the answer, and that process is manual and repeatable, so software can run it.

FinMoon is built on that idea. For each decision it runs a structured process modelled on how an expert works, personalised to your situation, and ends with a report.

Product principles I’d reuse anywhere

Encode the process and let the model fill it in. We design the workflow up front, the way an expert would approach the decision: which inputs to collect, which criteria to evaluate, which risks to check. Inside each step the language model does what it’s good at, which is reading dense documents, reasoning over the user’s specifics and explaining trade-offs in plain language. Splitting the work this way keeps the output consistent and makes it possible to review.

Most of the value comes from personal context. A credit card recommendation that doesn’t know your spending pattern is just a listicle. The more the system understands your situation (safely, and with your consent), the more useful each decision gets.

Design for the second decision. Because decisions interact, the product has to remember earlier ones, so your emergency fund setup informs the next insurance decision. That’s a data-model problem long before it’s an AI problem.

Get people to something useful before they sign up. Shared decision conversations work without a login, so someone can watch the full process on a real question before committing. Sign-up comes in when they want it personalised.

Google for Startups Hub at T-Hub

FinMoon was selected for the Google for Startups Hub at T-Hub, Hyderabad. For us it confirmed that the problem is real, and that building specifically for India (Indian products, tax rules and financial habits) sets FinMoon apart from tools that bolt on localisation later.

What building it has taught me

I’ve spent most of my career as an engineer and engineering leader inside other people’s products, which I wrote about in From developer to founder. Co-founding changed the questions I ask.

  • I used to ask whether we could build something. Now I ask whether we should, and in what order. The technical roadmap is a sequence of bets on what users will value.
  • Architecture decisions decide how long the money lasts. Model selection, hosting and caching are line items on a very short budget, so cost per useful answer is a product metric. I go into that side in Architecture before AI.
  • In finance, users stay only if they trust the answers. One confident wrong answer costs more than ten good ones earn, so evaluation, guardrails and explaining why the system recommends something have to ship with the first version.

We opened the beta to our first 200 users. If you make financial decisions in India and want to make them like an expert, try FinMoon. And if you’re building an AI product in a regulated, high-trust domain, I’m always happy to compare notes.