The architecture that makes AI stop inventing financial numbers

The core problem. Financial analysts need numbers that clients and portfolio managers can explicitly validate. Current LLM-based tools treat this as a semantic search problem, attaching sources after the fact. But LLMs do not actually read documents—they chunk, embed, and retrieve semirelevant information. This produces wrong sources, right sources with wrong numbers, or numbers without context (billion vs million, 2023 vs 2025). For mission-critical financial work, this is unacceptable.

The architecture. Kepler developed a system where the model plans the answer but never writes a number. Retrieval finds the correct source (using Turbopuffer's hybrid BM25 + approximate nearest neighbors search). Deterministic code—written by humans, not the LLM—performs all arithmetic. A post-processing filter strips any number the system cannot trace back to a verified citation. The invariant is simple: every number in Kepler is clickable and leads directly to the source document.

What makes it hard. XBRL data is wrong roughly 10% of the time. Analysts also work with internal documents from SharePoint, Box, OneDrive, and Outlook, not just SEC filings. Kepler must handle structured data (XBRL tags) and unstructured data (internal memos) within the same architecture. Tenant isolation is critical—one firm's desks must never see another's research. Every query is reauthorized in Postgres, and every filter is built incrementally.

The numbers that matter. When Kepler ran its filter across native Claude or ChatGPT output, it refused 10,981 citations from its evaluation set. Of those, 3,810 were flat-out wrong numbers. This is the hidden cost of naive RAG: hooking an LLM to a vector database is not enough for numerical accuracy. A third of all citations in Kepler's corpus are numbers the model did not write—they come from the deterministic compute layer.

The trade-off. Kepler's guarantees come with a cost. The platform is "structurally incapable of producing hallucinated numbers," but this also means it cannot be a creative thought partner. It is not useful for marketing content or speculative investment banking where numerical accuracy is secondary. The system can attempt citation repair—figuring out on the fly where a number came from—but this is where Kepler maintains its proprietary edge.

Where this leads. The speaker argues that the current generation of AI financial tools are limited by architecture, not by model capability. Fully managed agents can generate ideas from existing information but cannot produce a defensible discounted cash flow model. The future is "quantamental"—combining unique retrieval mechanisms with deterministic computation. This lets AI handle the work nobody wants to do while leaving creative analysis to humans.