AI Engineer

AI Can't Make Up a Number It Can't Look Up - Vinoo Ganesh, Kepler

Published 2026-11-07 · Duration 18:28

Summary

The core challenge in applying AI, particularly in finance, is not merely hallucination, but a structural lack of trust due to the inability to verify numbers and facts. The speaker details how traditional RAG platforms are insufficient because they fail to address the three sources of wrong numbers: retrieving the wrong document, extracting the wrong number from a correct document, or mathematical errors by the model. Kepler's solution is an architecturally different system that uses pre-processing to assign unique identifiers (pointers) to every fact, allowing the model to plan and point to these IDs rather than generating numbers itself. Verification is achieved through mandatory pointer dereferencing and post-processing checks, ensuring that every output fact is traceable to its source.

Download summary

Key takeaways

  1. The Core Problem: Structural Trust 4:10

    The fundamental issue with current AI systems is structural. Simply stating that models will 'get better' to eliminate hallucination is nonsensical. For high-stakes applications, the system must be designed so that the output can be verified, corrected, or refused.

  2. Three Sources of Error 6:57

    Wrong numbers originate from three sources: 1) Retrieving the wrong document; 2) Pulling the wrong number from a correct document; and 3) The model performing mathematical calculations (models predict tokens, they don't calculate).

  3. Kepler's Deterministic Architecture 10:17

    The system operates by assigning pointers to unique identifiers to every fact during a pre-processing step (unstructured to structured data store). The model acts as a 'smart analyst' that plans the solution and retrieves data via these pointers, preventing the structural generation of numbers.

  4. Advanced Search and Security 15:26

    To enable wide-then-narrow search, the system uses a combination of ANN + BM25 and cross-encoders. Security is paramount, requiring mechanisms to scope searches using OAuth tokens and prevent cross-tenant data leakage.

More digests on Generative AI

Technical details

  • Fact Extraction and Structuring 650s

    The system performs a giant pre-processing and contextualization step to convert unstructured data into a structured data store, assigning pointers to unique identifiers for every fact (e.g., a number in a 10K).

  • Model Interaction and Verification 730s

    The model is restricted from structurally generating numbers; it must use a mechanism of 'dereferencing' the assigned pointer to output a fact. This is enforced during a post-processing verification step.

  • Search Optimization 1000s

    The architecture allows for wide search (using ANN + BM25 to get a set of candidate documents) followed by narrowing (using reciprocal ranking and cross-encoders) to overcome the inherent limitations of LLMs in broad searches.

  • Security Scoping

    Critical security considerations include filtering the tenant, preventing caller impersonation, and ensuring searches are scoped using mechanisms that constrain the set of documents returned with each query.

Mentioned resources

Channel & topics

Watch on YouTube · Back to latest

This independent, AI-assisted summary is provided for commentary and informational purposes. It may contain errors or omit important context. Please watch the original video for the creator's complete presentation. Video, thumbnail, and related copyrights belong to their respective owners.