ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
NewsMaterial4 min readDisconfirmAI Editorial

Hrishi Olickel, Founder of Southbridge AI: No Universal Data Schema Exists; Structures Should Be Built Instantly to Support Decisions

Using self-reported data from an Ohio campaign finance entity resolution project at the AI Engineer conference, Hrishi Olickel argues that data transformation requires four modes (source, storage, decision, and model priors) that are only simultaneously known at the moment of decision, necessitating the instant construction of ontologies rather than pre-set universal schemas. He separates models into three roles—designer, executor, and optimizer—and uses cross-model disagreement for centralized review, noting that shared priors can still lead to consistent errors.

2026-10-13 00:00 GMT+8

A data pipeline stripped minus signs from loan amounts, turning a net $784,000 into $5.3 million in debt, with multiple models validating this incorrect result. Hrishi Olickel, founder of Southbridge AI, used this failure at the October 2026 AI Engineer conference to illustrate his core argument: the final bottleneck in data transformation—intelligence and judgment—is now cheaply accessible, but the true difficulty lies in not knowing what information means until a decision is made.

Olickel posits that every data record is constrained by four modes. Using temperature as an example: the source mode determines whether you have a regional phone reading, a desk sensor measurement, or a watch reading; the storage mode limits what can be recorded—device firmware may be omitted, and recording a timestamp to the second might take longer than writing the record itself; the decision mode determines which omissions are fatal—assessing comfort requires multiple measurements throughout the day and confirmation of the measurer, while weather mapping cares about location and device changes, and installing sensitive equipment requires excluding the person as noise. The fourth mode is the model's prior, its belief about how the world 'should be.' These four modes are only simultaneously known when the decision actually occurs. He cited the term 'customer': for billing, it is the payer; for support, the caller; for compliance, the entity the organization is responsible for. A unified customer box cannot resolve these distinctions before the question is asked.

Sign in / Sign up
ReadingResearchRadarInvestment framework

His alternative is a 'just-in-time ontology': building transformation structures when the decision is known, preserving their context and evidence for reuse. Operationally, he splits data work into three model roles. Designers decompose problems, write specifications, and authorize budgets, suited to slower, heavier reasoning models. Executors read PDFs, write database records, and check results within pipelines, suited to cheap, fast models with predictable defects. Optimizers start from a working run, try alternatives, debug, take notes, and adjust budgets to limit exposure, requiring reusable knowledge and stable execution environments. This separation explains why a new executor model could enter the Ohio pipeline without taking over its design—the system can test where the model helps while leaving specification and optimization to models suited for those tasks.

The Ohio campaign finance entity resolution project provided concrete figures. Olickel reported that the project was originally expected to cost about $10,000; after introducing a new decision model, the team completed all donor cases in ten days for under $800. The statewide job cost $795, with the first answer costing $128, while second opinions, disagreement resolution, and auditing totaled $545—checking costs were roughly four times the answering cost. On approximately 200 held-out families, a cheap reasoning model alone achieved results comparable to a comparison model at about 2% of its cost. A key finding was that a decision-first approach correctly answered core questions but missed nine household links because it confidently answered 'no' to each association. The team did not change models or thresholds but rewrote the household question to clarify 'what counts as sufficient evidence,' causing thirteen development questions to shift from no to yes. This demonstrated that definitions themselves can change results without changing the model.

However, Olickel explicitly marked the boundaries of this method. Cross-model disagreement can concentrate review—on the 200 held-out families, two models agreed on 194, none of which were wrong, with all observed errors falling among the six disagreements. Yet agreement is a fallible signal: in another benchmark, four model families produced identical error totals down to the cent because they shared a sign convention different from the answer key. Distillation may make seemingly independent models less so. Furthermore, all the above figures—from $10,000 to $795, from 194/200 agreement rates to 15,658 misclassifications—are self-reported results from Southbridge’s own projects, with no independent replication. His expectation that the next state’s audit budget would approach zero is a prediction, not a verified second deployment. His conclusion is conditional: any ontology is valid if its modes are labeled and context preserved; cheaper intelligence makes creating and adapting them more practical, and attached evidence makes reuse intelligible.

Sources

  • youtube.com↗
Continue researching →