Vinoo Ganesh, Kepler’s CEO and a former Citadel business-engineering leader, describes source-backed financial answers in this 18-minute AI Engineer talk.

Separate three ways numbers go wrong

  • Retrieval can select the wrong document or the wrong source for the firm’s question.
  • Extraction can select the wrong number inside the right document, such as the wrong reporting year.
  • Arithmetic can fail when the model generates an answer instead of calling calculation code.

Make facts addressable

  • Kepler preprocesses documents into structured facts with identifiers before answering questions.
  • The model plans which facts to use; code resolves their identifiers and performs calculations.
  • Clickable number chips expose the underlying evidence to the reader.
  • A final verification step checks the output and supports verified, corrected, or refused results.

Search broadly within enforced access rules

  • Combine approximate-neighbor search with BM25 keyword retrieval, then fuse rankings and apply a cross-encoder reranker.
  • Ganesh describes up to 200 initial candidates, narrowed to roughly 60 before the next ranking stage.
  • Enforce tenant identity and document access on direct reads as well as search.
  • Handle stale data and unpopulated indexes explicitly, alongside cross-firm data-leak risks.

Verify the meaning as well as the lookup

  • A valid identifier proves that a stored fact exists; its extraction, date, units, and relevance still need checking.
  • Ganesh later acknowledges that models can still generate unsupported numbers and need post-processing.
  • His production slide describes fewer refusals, without a narrated rate, sample size, or independent accuracy comparison.
  • Ask vendors to expose calculation sources and explain exactly what happens when a number has no supporting fact.

“You can’t just tack verification on at the end.” — Vinoo Ganesh, 1:18