Kepler CEO: models point to pre-extracted fact IDs instead of writing numbers — why Wall Street trusts it
AI Engineer · youtube · 2026-11-07
Kepler CEO and co-founder Vinoo Ganesh (formerly head of business engineering at Citadel) explains why Wall Street still doesn't trust AI and how to build systems it can.
Key points:
- Wrong numbers come from three causes: the wrong document, the wrong number from the right document, and models doing math. Lookups can't get it wrong; math and retrieval can.
- Trust means someone to blame — pure generative models can't offer that.
- Kepler's approach: code extracts every fact ahead of time and gives it an ID; the model plans and points to those IDs instead of writing numbers itself; anything it can't resolve gets corrected or refused.
Also covered: wide-then-narrow search, document-access scoping so firms never see each other's data, and the questions to ask any AI vendor about verification. His conclusion: finance is out of the search/pilot phase and into production.
More from coding & agent
- AI agent simulation of Pepys and the Great Fire loops: wake, look, sleep — repligate · 2026-11-07
- Devs bet on Flash-tier LLMs (~$1/M tokens) as agent workhorses, with Opus as orchestrator — bendee983 · 2026-11-07
- mitsuhiko: after repeated attempts, cloud dev environments still lose to his own infra — mitsuhiko · 2026-11-07
- mitsuhiko: coding in cloud sandboxes still loses to a well-provisioned local machine — mitsuhiko · 2026-11-07
- Uber Engineering Publishes Lessons on Securing AI Agents From Real-World Deployment — blaizedsouza · 2026-11-07
- Why parallel git worktrees break integration tests, and the 3-part isolation fix — Independent-Sir3234 · 2026-11-07