Kepler, a 2025 startup founded by ex-Palantir engineers Vinoo Ganesh and John McRaven, has launched Kepler Finance, a research platform that pairs Claude’s reasoning with deterministic infrastructure to give analysts auditable answers in plain English. According to Hacker News, the company built the system specifically to solve the trust problem at private equity firms, hedge funds, and investment banks, where every figure in a regulatory filing or deal pitch has to be verifiable against source documents.
The pitch is simple. Existing analytics tools can pull data but still need a human to verify it. Pure AI tools interpret freeform questions but generate the numbers themselves, which means they can hallucinate. Kepler splits the job in two: Claude handles interpretation and planning, deterministic code handles every calculation that has to be provably correct.
What Kepler shipped
In under three months, the team indexed an aggressive corpus to back the platform:
- 26M+ SEC filings
- 50M+ public documents
- 1M+ private documents
- 14,000+ companies across 27 global markets
The stack runs on AWS with Rust, Python, and containers for orchestration. Analysts type questions in English and get answers they can audit line by line.
How the system is wired
Kepler’s architecture is the interesting part. The team treats the language model as one stage in a pipeline rather than the whole brain. McRaven’s framing: “Prompt engineering optimizes a call while content engineering optimizes the system around it.”
That translates into several concrete pieces:
- Deterministic execution environments Claude invokes for any operation that needs to be exact, like computing a ratio or resolving a fiscal period.
- A proprietary ontology that maps financial concepts to precise definitions and formulas, customizable per use case.
- Security and access controls enforced at every step, governing which data sources each user can touch.
- Idempotent skills for recurring workflows like enterprise value calculations across complex capital structures (preferred shares, convertibles, minority interests) and segment revenue waterfall reconciliation.
- Multi-model routing: Opus 4.7 handles complex reasoning, intent decomposition, and ambiguity resolution. Sonnet 4.6 handles higher-throughput, more constrained stages.
- Specialized recall models, some built on Claude, some proprietary, hitting 94% accuracy on mapping financial statement labels to standardized taxonomy codes. Other models scored 38 to 46% on the same task.
Why Claude
The team benchmarked across frontier models. On simple queries, results were comparable. On long multi-step plans with interdependencies, every model except Claude started cutting corners or dropping constraints by the fourth or fifth step.
“On our workloads, Claude was the model that consistently held the plan together,” Ganesh says. “Other models would start strong and then quietly drop a constraint by step five.”
The behavior that mattered most wasn’t raw accuracy, though. When a term had two possible meanings, most models picked one and moved on. Claude stopped and asked. “That behavior matters more than any benchmark score,” Ganesh told the team. “One wrong assumption early in a financial analysis breaks everything downstream.”
Why this matters
Kepler is built around a simple thesis: in regulated industries, the model can’t be the whole system. The interesting part isn’t that Kepler uses Claude. It’s that the team explicitly engineered around the model rather than handing it the keys. Deterministic compute for math. Domain ontology for definitions. Hard boundaries on what to resolve versus what to escalate to a human. Automated evaluations that compare every prompt change and model upgrade against thousands of known-correct answers before anything ships to production.
That’s a template other regulated verticals (legal, healthcare, audit) will likely copy. The takeaway for builders: in domains where errors compound, the win comes from the scaffolding around the LLM, not from the LLM alone.
The Hacker News writeup doesn’t mention pricing or general availability terms, so firms interested in access will want to check Kepler directly. Full details at the original source.