Resect launches with $25M to reduce hallucinations in AI models
SiliconANGLE Kyt Dotson
Resect AI raised $25 million to catch AI hallucinations as they happen. It’s betting enterprises want models that can be audited, not just told to behave.
Based on reporting by SiliconANGLE, Kyt Dotson — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Seattle startup Resect AI has raised $25 million in early funding to build what it calls an accountability layer for enterprise AI. The target is familiar and embarrassing: hallucinations, where a model confidently spits out something false or made up.
Chief executive Kevin Owens says AI has been trusted too early, and the usual warning labels don’t cut it when companies need governance and compliance. Resect’s pitch is that models should be anchored in truth if enterprises are going to rely on them at scale. That’s a stronger claim than “safer AI.” It’s a shot at the whole black-box problem.
The company says its work started with building its own models for high factuality, then shifted once it figured out the same training approach could be applied to open models including DeepSeek, Qwen and Llama. Instead of chasing a better model from scratch, Resect is now building tools that sit inside model architecture, watch for misbehavior, and redirect output before a hallucination lands.
Chief AI officer Tim Walton says the team has spent a lot of time studying how models choose answers and where they fail. The company’s enterprise audit layer is called NeuroWave Product Suite, and Resect describes it as a kind of polygraph for neural networks. The name is dramatic, but the goal is plain enough: observe, detect, interpret, audit and modify model behavior.
Resect currently has two open models on HuggingFace, a 0.6-billion-parameter Veritas fact checker and an 8-billion-parameter model, both based on Qwen3 and both non-thinking, meaning they don’t use a chain-of-thought layer. On LLM-AggreFact, the 0.6B model scored 72.3%, which the company says is a 7.4% improvement over Qwen3. The GitHub repository for its open-source tools exists, but it wasn’t populated at publication time. Resect didn’t name its backers, only saying the money came from private equity investors and will go to research, go-to-market work and hiring around Seattle and Portland.
My take — AI-written commentary, not fact-checked reporting
This is the right fight, and a nice reminder that “AI safety” is often just procurement with better branding. Enterprises don’t need another model that can talk confidently; they need one that can be checked when it starts making things up. Also, “polygraph for neural networks” is exactly the kind of phrase that makes the hype crowd clap and the compliance team reach for coffee.
Read more about this at: SiliconANGLE