Sign InOpen Brain
AI EngineerVideoSource Linked

How Kepler Built Verifiable AI for Financial Services — Vinoo Ganesh

Kepler’s pattern keeps LLMs away from numeric mutation: agents choose sources and steps, while deterministic systems persist, calculate, and reject unverifiable values.

AI Engineer · Jul 29, 2026
Open Source Open MarkdownOpen JSON
Source Summary

Kepler separates citation from verification: a source shows where information came from, while verification deterministically checks that a value is correct. Its extraction model reached **94% accuracy**, but the remaining errors were still unacceptable for financial work.

Practical Implication

For high-stakes agents, let the model plan and select references, then hand numeric I/O and calculations to deterministic tools. Keep an **atomic provenance ledger** and remove any value that cannot pass an independent check.

Agent-Ready Context
Kepler separates citation from verification: a source shows where information came from, while verification deterministically checks that a value is correct. Its extraction model reached **94% accuracy**, but the remaining errors were still unacceptable for financial work.

For high-stakes agents, let the model plan and select references, then hand numeric I/O and calculations to deterministic tools. Keep an **atomic provenance ledger** and remove any value that cannot pass an independent check.

Verification does not establish one universal conclusion. It proves that the replayable path followed an organization’s own rules, and the talk leaves personalization of those rules and ontologies as the next major problem.
Connected Context · Feed7 Judgment

This sharpens reliability for financial agents by separating traceability from correctness: citations are insufficient unless numeric inputs and calculations can be independently replayed and invalid values removed. It reinforces deterministic tool boundaries and atomic provenance, while narrowing formal verification claims to compliance with an organization’s encoded rules rather than universal truth. The 94% extraction result explains why residual model error remains operationally unacceptable.

Build for the Memo, Not the Demo — Shawn Chan, China Resources HoldingsClaim-level provenance and diligence controls are strengthened here by an atomic replay ledger that independently verifies numeric values instead of merely showing their sources.Semantic Blindness: 500,000 Sensors Confused an LLM - Raahul Singh & Vanč Levstik, PhaidraBoth assign ambiguous planning to the model and exact resolution to deterministic machinery; Kepler applies the same boundary specifically to financial numeric I/O and calculations.CausalForge: A Formally Grounded, Self-Improving Agentic Framework for Automated Research in Causal InferenceBoth limit what verification proves: CausalForge checks derivation without guaranteeing the intended claim, while Kepler checks a replayable path against organization-specific rules without establishing a universal conclusion.Stop AI Agent Hallucinations: 5 Techniques + Production Patterns - Elizabeth Fuentes, AWSAWS’s validation and pre-call rules provide broad runtime controls; Kepler narrows them into a high-stakes policy that excludes every financial value lacking an independent deterministic check.
Context Map
agentdata#harness-engineering#tool-use#agent-reliability
Uncertainty
Verification does not establish one universal conclusion. It proves that the replayable path followed an organization’s own rules, and the talk leaves personalization of those rules and ontologies as the next major problem.