Resect AI is developing an accountability layer for enterprise AI to detect and reduce hallucinations at runtime. The startup is building an open-source offering that allows users to observe, interpret, audit, and modify the internal behavior of large language models, effectively opening the 'black box' of AI internals.
The company's technology is based on research into how models think and why they produce specific answers. Resect initially developed its own high-factuality models and discovered that its training and reinforcement processes could be applied to existing open models such as Llama, Qwen, and DeepSeek.
These internal visibility tools will be integrated into an enterprise-level audit tool called the NeuroWave Product Suite, which the company describes as a 'polygraph for neural networks'. This suite is designed to intercept and surgically fix model misbehavior before hallucinations occur.
Currently, the startup has released two models on HuggingFace: a 600 million-parameter '0.6b' Veritas fact checker and an 8 billion-parameter model. Both are based on the Qwen3 architecture and operate without a chain-of-thought layer. The 0.6B model showed a 7.4% performance improvement over Qwen3 on the LLM-AggreFact benchmark, scoring an average of 72.3%.
The company also plans to offer open-source tools via a GitHub repository, although the repository is not yet populated. The funding will be used for research and development, go-to-market initiatives, and hiring local talent in the Seattle and Portland regions.
