Do Not Let Models Invent Numbers - Vinoo Ganesh, Kepler
LLMs hallucinate financial numbers because they chunk documents semantically, not read them literally
“The same thing that makes them so creative and cool makes them very difficult to use in critical situations.”
Kepler CEO Vinoo Ganesh argues that adding citations after LLM generation is insufficient for financial use cases because LLMs perform semantic chunking rather than true document reading, leading to numbers sourced from unreliable third parties. Kepler uses TurboPuffer as part of an architecture validated by Anthropic to enforce number provenance before generation. The talk is a practitioner-level case study on grounding LLM outputs in verifiable financial data.