Seattle-based artificial intelligence research startup Resect AI announced Thursday it has raised $25 million in early funding to build an accountability layer for enterprise AI by capturing and reducing hallucinations at runtime.

Hallucinations, or confabulations, happen when an AI model replies with a false or fabricated response and presents it with high confidence.

At their core, models operate by predicting patterns based on training, not “remembering.” This means that gaps in data or a lack of real information lead them to guess. Additionally, many models are designed to appear helpful during refinement and post-training; instead of generating an “I don’t know” reply, they can confidently give wrong answers.

“AI has prematurely been put in a position of trust,” said Chief Executive Kevin Owens. “Adding labels such as ‘use at your own risk’ flies in the face of proper governance or compliance. AI must be anchored in truth to be widely adopted across the enterprise.”

Owens added that Resect aims to bring accountability and transparency to AI models, which until now have been black boxes. Resect AI is building an open-source offering for enterprise products that opens up that black box to look inside large language models to observe, detect, interpret, audit and modify that behavior.

“Many argue that understanding the black-box internals of LLMs is out of reach, but we fundamentally disagree,” said Chief AI Officer Tim Walton. “We’ve spent an extensive amount of time and resources researching how models think, and what causes them to choose the answers that they do.”

Resect’s journey started with rolling its own AI models capable of producing high factuality. According to the company, it built its own model to control data, developed a training process, behavior and reinforcement training, which it discovered could be applied to existing open models including DeepSeek, Qwen and Llama.

After that discovery, the task became less about building a better model and turned toward building a suite of tools that could work within model architecture itself to intercept misbehavior and redirect it before hallucinations happen. That means understanding why AI behaves the way that it does internally.

The company said these internal visibility tools will form the basis of an enterprise-level audit tool called the NeuroWave Product Suite, described as a polygraph for neural networks.

“We’ve spent an extensive amount of time and resources researching how models think, and what causes them to choose the answers that they do,” Walton explained. “Through this process, we’ve developed technology that observes exactly when and how models fail, and surgically fixes them.”

The surgical language is especially fitting. The name of the company, “Resect,” is a medical verb that means to cut out or remove part of an organ, tissue or bone – a surgical operation that involves the careful cutting away of damaged tissue while leaving healthy tissue untouched.

Currently, the company offers two models on HuggingFace, the model repository: a 600 million-parameter “0.6b” Veritas fact checker and an 8 billion-parameter model. Both are based on the Qwen3 architecture and operate non-thinking, meaning they do not have an underlying chain-of-thought layer. According to the model cards, the performance of the models was evaluated on LLM-AggreFact; the overall performance of the 0.6B model scored an average of 72.3% compared with Qwen3, an improvement of 7.4%.

The company also links to a GitHub repository where it will offer open-source tools, but it’s not populated yet.

Resect didn’t announce the names of its backers, but said the funding came from private equity investors. The company said the funding will go toward research and development, go-to-market initiatives and local talent hiring in the greater Seattle and Portland regions.

Image: SiliconANGLE/Microsoft Designer