Kepler’s Blueprint: Making AI Outputs Auditable in Finance with Claude

Financial analysts crave speed. They demand proof. Kepler delivers both.

Founded in 2025 by Vinoo Ganesh and John McRaven—veterans from Palantir’s data systems for defense and finance—the startup tackled a core industry pain. After interviewing 147 firms from private equity to investment banks, the founders heard one refrain: AI promises efficiency, but outputs lack traceability. “How am I supposed to trust something I can’t audit?” one managing director asked, as detailed in Claude’s blog post.

Kepler Finance indexes over 26 million SEC filings, 50 million public documents, 1 million private ones, covering 14,000 companies in 27 markets. All in under three months. Analysts query in plain English. Answers appear instantly—each figure linked to its source filing, page, line item.

But. Finance forbids errors. Regulated reports demand audit trails. Traditional tools extract data. Analysts verify manually. AI interprets queries and computes, but bundles reasoning with execution. Hallucinations creep in.

Kepler splits the load. Claude handles reasoning: query decomposition, ambiguity resolution, plan creation. Deterministic layers execute: retrieval, computation, verification. Rust, Python, AWS containers enforce precision. Every ratio, fiscal adjustment—provable.

“On our workloads, Claude was the model that consistently held the plan together,” Ganesh says in the Claude blog. Others falter by step five, dropping constraints. Claude flags uncertainty. “That behavior matters more than any benchmark score,” he adds. One bad assumption cascades.

Engineering Claude for Multi-Step Precision

Kepler feeds Claude structured inputs: proprietary ontology maps terms to formulas. Security gates data access. Custom skills tackle repeats—like enterprise value across preferred shares, convertibles.

Pipeline stages match models. Claude Opus 4.7 reasons complexly: intent breakdown, plan structuring. Sonnet 4.6 speeds constrained tasks. Kepler’s fine-tuned models hit 94% on label-to-taxonomy mapping; others lag at 38-46%.

Tests run thousands of cases per change. Automated pipelines check plans and results against ground truth. New Anthropic models benchmark in hours—pinpointing gains, regressions, prompt tweaks needed.

Idempotent design shines. Same input, same output. Every time.

Scale follows. Claude interprets vast unstructured data, reconciles terminology shifts. Retrieval pulls verified figures. Results export to Excel templates. One click traces lineage.

Small team, big wins. Modular architecture ships features in weeks, not months. Compliance baked in: SOC 2 Type II certified, ISO 27001 pending. Audit logs, siloed environments, full provenance from launch.

Finance tested the build. Dense data. Overloaded terms. Zero-error math. Survived? Yes. Now domain-agnostic. Healthcare trials, legal precedents—same pattern: Claude reasons, infrastructure proves.

“Kepler Finance is our first product,” Ganesh notes. “It won’t be the last.”

Industry Momentum Builds on Verifiable Foundations

Anthropic’s broader push amplifies this. Claude for Financial Services launched July 15, 2025, topping Vals AI’s Finance Agent benchmark, as per Anthropic’s announcement. Integrates Snowflake, S&P Global, Morningstar. Excel plug-in builds DCFs, coverage reports.

October updates added Agent Skills: discounted cash flows, market connectors. Firms like BCIM dashboard comps from S&P, FactSet. Norges Bank queries 9,000 holdings daily.

FactSet integrated Claude February 28, 2026, deepening workflows amid AI competition, reports Yahoo Finance. (Note: separate Kepler Cheuvreux partnership mentioned, distinct from startup.)

Hacker News buzzed recently, with 32 points on the Claude post, sparking systems debates, via Hacker News.

Kepler’s CTO McRaven stresses: “Prompt engineering optimizes a call while content engineering optimizes the system around it.” Finance pros verify instinctively. Provenance isn’t bolted on—it’s foundational.

Best practices emerge. Assign Claude interpretation, not retrieval or math. Match models to tasks. Test rigorously. Build audit-ready from day one.

AI in finance advances. Not through smarter models alone. Through verifiable execution. Kepler proves it works.