Financial services firms have long wanted to use AI for research but have been held back by a fundamental problem: they cannot trust the output. Kepler, a startup founded in 2025 by former Palantir engineers Vinoo Ganesh and John McRaven, has built a platform that separates AI reasoning from deterministic computation, making every answer auditable.
Overview
Kepler Finance is a research platform for financial services. Analysts ask questions in plain English and receive answers that can be traced back to the exact line item in the source document. The platform uses Anthropic's Claude as the reasoning and interpretation layer, but the actual computation — calculating ratios, resolving fiscal periods, pulling figures — happens in deterministic execution environments that Claude can invoke but cannot override.
The problem
Before founding Kepler, Ganesh and McRaven spoke with 147 financial firms — private equity, hedge funds, investment banks. Nearly all said the same thing: they wanted to use AI, but they could not trust the output. As one managing director put it: "How am I supposed to trust something I can't audit?"
Traditional analytics tools can pull data but cannot interpret freeform questions or decompose them into steps. AI systems can interpret, but they handle interpretation and computation in the same step, so the numbers they produce are generated by the model — which can make mistakes.
How it works
Kepler's architecture is a multi-stage pipeline. Different Claude models handle different stages:
- Opus 4.7 for complex reasoning: decomposing intent, resolving ambiguity, producing structured execution plans.
- Sonnet 4.6 for higher-throughput stages where tasks are more constrained.
The team benchmarked all frontier models. On straightforward queries, performance was comparable. But on long, multi-step plans with interdependencies, all models except Claude started taking shortcuts or losing track of constraints by the fourth or fifth step. "On our workloads, Claude was the model that consistently held the plan together," Ganesh says.
The clearest difference was how each model handled uncertainty. When one term could have two meanings, most models picked one and kept going. Claude stopped and asked the analyst to decide.
Kepler also built a proprietary ontology that maps financial concepts to precise definitions and formulas, customizable per use case. They trained their own specialized models for recall (some use Claude as the foundation, some are proprietary), scoring 94% accuracy on tasks like mapping financial statement labels to standardized taxonomy codes, compared with 38-46% for other models.
Scale and verification
Kepler Finance has indexed more than 26 million SEC filings across 14,000+ companies, 50 million public documents, and