Just-in-Time Ontologies: when multiple AI models agreed on turning $784K debt into $5.3M

AI Engineer · youtube · 2026-10-13

In an AI Engineer talk, Hrishi Olickel (Southbridge AI) explains why LLM-based data transformation needs more than a universal schema, using a pipeline failure that stripped minus signs from loan amounts — turning net $784,000 into $5.3 million of debt, with multiple models agreeing on the wrong answer.

Key points:

"Customer" means different things to billing, support and compliance; useful structure depends on source, storage constraints, the decision at hand, and model priors
His alternative: just-in-time ontologies — build the transformation when the question is known, and preserve context and evidence for reuse
An Ohio campaign finance entity-resolution project illustrates separating models that design the pipeline, perform its jobs, and optimize the system
Definitions, held-out cases and disagreement checks matter as much as model selection; cross-family model comparison focuses review on disagreements, but shared assumptions can still yield the same error
Self-optimization loops need a stable execution surface, bounded budgets, reusable methods, and a defined way to fail
Small pilots establish baselines; independently checked work segments make revisions easy to inspect and revert

Original post →

Just-in-Time Ontologies: when multiple AI models agreed on turning $784K debt into $5.3M

More from coding & agent