TechCrunch: Probably raises $9M for AI with deterministic validation

Probably, founded by Peter Elias, announced a $9 million seed round led by Andreessen Horowitz on June 16, 2026. The company’s premise is that most AI reliability problems are engineering problems rather than model problems, and that the right solution is a validation architecture around the model rather than a more capable model.

Elias describes the approach as a “data science mech suit”: LLM outputs pass through a deterministic validator before reaching the user. If the validator cannot confirm the result against source data, the output is flagged or blocked. The architecture makes it possible to run on a model Elias describes as “four classes weaker than frontier models,” because model capability is partially compensated by the quality of the validation wrapper. Smaller models can run on local hardware rather than data center infrastructure, which substantially reduces inference costs — a meaningful advantage as organizations face growing pressure on AI operating budgets.

The current product focuses on data science queries: users ask questions about complex datasets and receive cited, auditable answers. Elias claims accuracy at 99.99%, in line with what traditional deterministic software achieves, while preserving the flexibility of natural language interfaces. Planned expansions include accounting and medical services — fields where factual errors are liability issues rather than product quality problems.

For product managers building AI features in regulated or precision-sensitive environments, the architecture Probably describes is a concrete design pattern to understand. The relevant question shifts from model selection to what validation sits downstream of the model. Teams that have accepted a tolerable error rate in their LLM outputs may find that a validation-wrapper engineering approach changes both the cost structure and the risk profile of their AI feature — particularly if they are currently using frontier models primarily because they need the accuracy, rather than because they need the full capability range those models offer.