Probably raised $9M to fix LLM hallucinations by betting on smaller models
probably raised $9m from andreessen horowitz to build a system that catches llm errors using deterministic validators, allowing it to run on local hardware instead of expensive data centers.
startup probably raised $9m in seed funding led by andreessen horowitz to address persistent llm hallucination issues through a novel 'harness' engineering approach. instead of relying on smarter models, the company uses a deterministic validator that checks answers against datasets, bouncing errors back to the model until the output is verified, achieving 99.99% accuracy. this architecture allows the system to run on models significantly smaller than frontier models, enabling operation on local desktop hardware rather than expensive data center gpus and cutting token costs. the company is launching first with a data science tool but plans to expand the engine to accounting and medical services.