Our Research
Original research from the engineers building Kepler: how AI can read everything, know a firm's world and prove every number.
Engineering Research
Better language models won't fix hallucination. They'll make it quieter.
Better models make wrong numbers rarer and harder to see. The fix is a provenance layer where the model names facts by reference and never retypes the digits.
Research noteStop Paying to Forget
Most document AI pays to reread documents for every question. We read each one once, keep what we read, and answer the next question with ordinary software.
Research paperReclassified to Conform
194 of 200 companies reshaped a financial statement in ways XBRL tags can't follow. Our paper recovers what changed from the filings' own arithmetic.
Financial Research
Anthropic S-1 Breakdown
Anthropic's S-1: $16.3B of H1 2026 revenue (up 1,300%), a Q2 operating profit, and why the $80.7B loss is mostly non-cash. Research note and Excel model.
Oura S-1 Breakdown
Inside Oura's S-1: revenue up 123% to $908M in FY2025, 5M paid members, positive free cash flow. Research note plus a model traced to the filing.
From the Blog
All postsBuild vs Buy AI in Finance
Buy what every firm needs, then build your edge. Where the build vs buy line sits for hedge funds, private equity, credit, allocators and banks.
Deterministic AI: What It Means, Where It Breaks, and Where It Belongs
Enterprise AI vendors all promise determinism. Here's what the word actually covers, why everyone is talking about it, and which parts of an AI system should be deterministic.
The Compiler That Never Says No
English is now a programming language whose compiler runs unclear instructions as readily as clear ones. Clear writing is the scarcest skill in technical work.
The platform behind the research
Everything here runs on Kepler, the same platform teams use on their own data, with every number traced to its source.