
Resect AI builds enterprise software that detects and corrects hallucinations inside large language models.
Resect AI sells software that reads the internal state of a running large language model rather than only grading its finished text. The NeuroWave suite observes activations, flags claims the model is not internally grounded in, keeps an audit record of each detection, and can rewrite a response before delivery.
Alongside the commercial suite the company publishes Veritas, a pair of openly licensed models at 8B and 0.6B parameters fine-tuned for factual grounding. Teams that prefer to self-host can run those weights directly instead of routing traffic through a hosted service.
Resect AI reaches the market at a moment when regulators and enterprise buyers are asking model operators to show their work, and when unverified generated text carries direct legal and reputational cost in publishing, finance, healthcare, research and education. That demand is what the company positions its audit trail against.
The category is contested. Model vendors are adding their own grounding and citation features, retrieval-augmented generation vendors sell an adjacent answer to the same problem, and an interpretability supplier must keep pace with each new model architecture it wants to instrument.
Most hallucination tooling grades a model answer after generation, comparing it against retrieved documents or a second model. Resect AI instead instruments the model itself, which lets it explain why a specific claim was produced and intervene on the generation rather than only score it.
Publishing the Veritas weights openly gives the company a distribution channel that its closed competitors lack, letting engineers evaluate the grounding research before any commercial conversation. Its buyers in publishing, finance, healthcare, research and education need exactly the written audit trail that an activation-level record produces.