The gap we kept running into

At Flipkart, I helped scale a BNPL product to 10 million users and a $1B loan book. Our in-house software originated those loans in seconds, literally. But each of those loans in the book was financed against borrowing lines, eligibility tests, concentration limits and monthly reporting pack. This part ran entirely on Excel by humans.

Every month, there would be a relay race between Finance, Operations and Business teams to make it to the finish line. There would be something off in each cycle and fixing it would take massive time wasted in back-forth between multiple teams. The regulator would drop a penalty and sometimes the Lender would just write us off for any line extensions.

Before that, I spent eight years at Goldman Sachs and UBS in M&A and debt capital markets, on the side of the table that writes these facility agreements. The legal documents are long and full of conditional clauses. Every new amendment silently changed the math of the entire operations. Most of the funds still carry that math in templates that someone has to rebuild by hand when the deal changes or worse - when the staff quits.

Why this is hard for generic LLM, and why we built a vertical harnAt Flipkart, I helped scale a BNPL product to 10 million users and a $1B loan book. Our in-house software originated those loans in seconds, literally. But each of those loans in the book was financed against borrowing lines, eligibility tests, concentration limits and monthly reporting pack. This part ran entirely on Excel by humans.ess

Large Language models like Claude and ChatGPT are good at reading documents. But they are not good at giving you the same answer twice. Fund operations has zero tolerance for that. A borrowing base that is 98% correct is wrong. Moreover, the token cost burnt to skim through 250-page legal documents every time does not justify the ROI of the task. We ran a benchmark of these tasks across multiple LLM models - read the results here.

So we don't ask a model to guess the number. We structure the data in a semantic layer along with domain specific tool calls that our clients LLM use to run their operations. We customize these tool calls for our client environment and use cases. That way the institutional knowledge and the self-learning loops stay with our client. We wrote about our security and certification in a separate post here.

The result is tasks that used to take weeks by multiple teams get finished in minutes. It gets reviewed by a human who has moved from being a Preparer to a Reviewer.

What YC changes

Mostly pace. YC helps you move mountains in minutes. In just one month, the velocity of company building has grown exponentially. Be it in go-to-market or product development or simply decision making. YC helps you in cutting the noise and laser sharp focus on things that moves the needle for your company. "Ambition" gets a whole new meaning at YC - we scaled up in our own ambitions and we are excited for the future.

If this sounds like your month-end

If you run operations, finance, treasury or portfolio management at a private credit fund, I would like to hear how it works out for you. Even if you're nowhere near buying anything, I would love to lean in to your best practices. I would love to grab coffee and trade notes.