‘We use prompts that have been tuned through GEPA, so the LLM isn’t freelancing’: Intuit Credit Karma’s Gurpreet Singh on pairing financial modeling with Gen AI

Financial firms that treat debt, refunds, and paychecks as separate products, then layer on generic AI chatbots, still leave customers to connect the dots on their biggest financial decisions.

Intuit Credit Karma takes a different approach. Rather than offering a general-purpose chatbot, it has built AI assistants for specific moments like paying down debt, managing tax refunds, and allocating paychecks. 

The Debt Assistant, Refund Assistant, and Paycheck Assistant, each help consumers navigate a specific financial decision. Using a shared financial context for each consumer, the assistants generate personalized recommendations and explain the reasoning behind them, helping consumers understand both the recommended action and the logic behind it.

To understand how this works in practice, Tearsheet spoke with Gurpreet Singh, Head of Product at Intuit Credit Karma, who explains the architecture behind the company’s AI strategy and why getting the underlying technological foundation right matters more than building fast.

Q: Debt Assistant pairs a financial optimization engine with an LLM. How do those two systems work together, and why not let the LLM handle recommendations on its own?

Gurpreet Singh, Intuit Credit Karma: The architecture is split deliberately because when you’re helping someone consolidate their debt, you cannot afford to get the numbers wrong – the stakes are too high. 

Debt Assistant works in two distinct stages, orchestrated together. 

First, we have a deterministic financial optimization engine built on event-based simulations and linear optimization. The engine ingests a member’s debt profile, including open tradelines, balances, interest rates, monthly payments, and their stated goals, whether they’re trying to minimize total interest costs or they need to lower their monthly payments to free up additional cash. From there, it evaluates different debt consolidation scenarios to solve for the mathematically optimal path, guaranteeing the savings number a member sees in their plan comes from an amortization formula, not an LLM’s best guess. 

The second stage is powered by generative AI (GenAI). 

…