AI startup Probably has raised $9 million in seed funding from Andreessen Horowitz to advance a new approach to more dependable language models. The company is focused on reducing hallucinations and factual slips before they reach users.
Founded by Peter Elias, the company aims for near-deterministic levels of accuracy in AI-powered workflows. Its first product is a data science tool designed to deliver fast answers from complex datasets, while also showing citations and a clear audit trail for each result.
To improve reliability, Probably combines an LLM with a deterministic validator system. If an output does not align with the underlying dataset, it is sent back for correction. The model is also trained to work with this validation layer, creating a tighter loop between generation and verification.
Elias says this method can reduce the need for larger, more expensive models by making the surrounding system more precise. In practice, the tool can run on smaller models and even local hardware, helping lower compute and token costs.
While the first use case centers on data science, the same architecture could support other precision-sensitive fields such as accounting and medical services. The broader idea is to make AI more trustworthy by designing systems that minimize ambiguity from the start.
This approach points toward a future where AI tools may become not only smarter, but also far more reliable across everyday professional use.
