Probably Raises $9M to Build AI Systems That Catch Their Own Errors

The Validation Layer That Changes AI Economics

Probably, a startup backed by Andreessen Horowitz with $9 million in seed funding, is rethinking how AI systems should be built to achieve the reliability standards of deterministic software. According to TechCrunch AI, the company’s foundational insight is that rigorous engineering around an LLM—not the model itself—is what delivers accuracy at scale.

Rather than relying on ever-larger models to reduce errors, Probably pairs a modestly-sized language model with what founder Peter Elias calls a “data science mech suit”: a deterministic validator that checks every LLM output against ground truth before it reaches a user. If an answer doesn’t match the underlying dataset, the system rejects it and iterates. The LLM itself is trained to work within this validation framework, optimizing for the constraints the validator imposes.

The result is a system that achieves 99.99% accuracy—a threshold common in traditional software but exceedingly rare in generative AI. According to TechCrunch AI, this level of precision is particularly valuable in “precision-sensitive use cases” like accounting, medical services, or data analytics, where hallucinations can be costly.

Why Smaller Models Become Possible

The efficiency gains are substantial. According to TechCrunch AI, Elias disclosed that Probably’s current product runs on a model “four classes weaker than the frontier models,” meaning it can execute on local hardware rather than cloud infrastructure. This shift has immediate cost implications: token expenses, which have become a significant concern as customers reassess AI budgets, drop dramatically when inference happens on commodity CPUs instead of high-end GPUs or distributed clusters.

The mechanism is counterintuitive but elegant. By refining context and removing ambiguity through the validator harness, the language model doesn’t need to work as hard. As Elias explains to TechCrunch AI, “If you can refine the context enough, the model does not have to work very hard to do the right thing. Basically, it’s an exercise in reducing ambiguity.”

Why This Matters

Probably’s approach exposes a structural misalignment between how large AI labs build systems and what enterprises actually need. According to TechCrunch AI, Elias notes that frontier AI companies are “incentivized not to” pursue this path, since higher error rates drive additional token consumption and revenue. For customers, Probably’s validator-first architecture offers a rare combination: lower operational cost (smaller models), faster inference (local execution), and higher reliability (deterministic validation).

The startup’s initial product focuses on data science tools with citations and audit trails, but the architecture generalizes. If Probably can execute on its vision across accounting, healthcare, and other regulated domains, it positions a new category of AI reliability layers—systems that work with smaller models rather than demanding larger ones—as the default for mission-critical applications. This could accelerate the shift away from the “bigger is more reliable” assumption that has dominated AI engineering since 2022.