Vinoo Ganesh, CEO of Kepler, explains the company’s financial-research architecture in a 21-minute AI Engineer talk hosted by turbopuffer.

Give each number a precise meaning

  • A financial fact needs a value, unit, scale and reporting period, plus the company and document it describes.
  • A filing published in 2025 can contain a 2023 revenue figure. Its table heading may supply the currency and scale.
  • Attaching a source after drafting leaves room for a relevant document to support the wrong figure.
  • Kepler assigns facts stable identifiers and stores their context in a citation registry.

Separate planning, retrieval and calculation

  1. The model plans which filings, disclosures and adjustments are needed to answer the question.
  2. Retrieval combines meaning-based search with keyword matching on turbopuffer, followed by re-ranking and source-trust checks.
  3. Human-written calculation code computes financial results and registers their supporting inputs.
  4. A post-processing filter resolves fact references and removes unsupported numbers before display.

Enforce permissions and inspect the evidence

  • Ganesh says document hits are reauthorized in Postgres, with scoped access tokens and filters carried through the retrieval process.
  • Public filings and private research need the same evidence handling, with access checks that keep desks and clients separated.
  • Clicking a number exposes its source and, for derived figures, the calculation used to produce it.
  • Some unmatched citations undergo repair. Ganesh leaves the repair policy partly undisclosed, so its acceptance criteria remain unclear.

Understand what the guarantees cover

  • Ganesh reports thousands of refused citations and wrong numbers in Kepler’s tests of ordinary document-backed model answers.
  • The talk supplies no complete evaluation set or denominator, so those counts cannot establish a general model error rate.
  • Rejecting unsupported figures trades answer coverage and speculative freedom for inspectable evidence.
  • Caveat: these are the vendor’s architecture and reliability claims. Source extraction, financial definitions, permission checks and citation repair still need testing.

“The model can still hallucinate and it definitively does.” — Vinoo Ganesh, 18:27