Resect AI raises $25M to inspect and modify enterprise LLM behavior

Why it matters

Resect AI is betting that enterprise AI spending will move toward tools that can explain and alter model behavior, with $25M buying time to prove that claim in production.

Kevin Owens, Tim Walton, Tyler Gerber and Tommy Lofgren took Resect AI out of stealth on September 3rd with $25 million to build tools that inspect large language models from the inside and intervene when they begin producing unreliable answers.

Resect AI described the financing in its September 3rd announcement as funding from private-equity investors. It did not name the backers, label the financing as a seed or growth round, or disclose a valuation. Resect AI said the money will fund research, sales and hiring around Seattle and Portland.

Owens, Resect AI's founder, CEO and chairman, previously worked with Gerber and Lofgren on Audiience, an advertising and data-science venture built around operating without third-party cookies. The three have reunited with Walton, Resect AI's chief artificial intelligence officer, around a harder technical problem: understanding why a neural network produces a particular answer and changing its behavior before a bad answer reaches the user.

That ambition makes the financing meaningful. Resect AI is entering a crowded market for AI evaluation and observability, yet Owens is placing the company deeper in the model stack than most enterprise monitoring products. Resect AI says its technology can inspect activity during inference, identify emerging hallucinations, modify model behavior and preserve an audit record of what happened.

A model-level bet

Resect AI's origin story began with the founders trying to train a more factual language model. According to its website, the group built its own model to control the data, training process and resulting behavior, then applied its reinforcement-training methods to open models including DeepSeek, Qwen and Llama.

The work pushed the founders toward interpretability. External evaluation could show that a model failed, but it could not necessarily explain the internal sequence that produced the failure. Resect AI now wants to sell that internal visibility as the NeuroWave Product Suite, which it describes as a polygraph for neural networks.

The planned suite includes real-time hallucination detection, behavioral modification, vLLM integration, audit logs and controls for AI agents. Resect AI says the tools are intended for publishing, finance, healthcare, research and education, fields where a fluent fabrication can become a compliance incident or an expensive operational error.

Resect AI says its early model work began with an effort to build a system that could be trusted. That thesis is more demanding than conventional output filtering. Direct intervention inside a running model must identify a genuine failure quickly, avoid blocking accurate responses and work across model families whose internal behavior differs.

Open source before the enterprise rollout

Resect AI has published 0.6-billion and 8-billion parameter Veritas fact-checking models through its Hugging Face organization. Both are based on Qwen3 and carry Apache 2.0 licenses.

On the 0.6-billion parameter model card, Resect AI reports average balanced accuracy of 72.30% on LLM-AggreFact, compared with 64.93% for the underlying Qwen3 model in non-thinking mode. That is a 7.37 percentage-point gain in Resect AI's published testing. The model card says the benchmark used an unseen test set, while also noting that Resect AI had submitted a pull request to add support for the model's operating mode to the MiniCheck library.

Those models provide an inspectable artifact behind the launch. The broader commercial claims remain tied to Resect AI's own product descriptions. Its enterprise suite is offered through a waitlist and demo request, indicating that September 3rd marked the public start of its go-to-market effort rather than broad availability.

Resect AI's website also directs developers to its GitHub organization for an open-source product. As of September 4th, that organization displayed no public repositories. The Hugging Face releases are currently the clearest public implementation of Resect AI's factuality work.

$25M with the backers offstage

Resect AI's financing gives Owens and his co-founders room to build a research-heavy product and recruit outside the largest AI hubs. GeekWire reported that Resect AI has about 30 employees, with four working from its Washougal, Washington, headquarters and others distributed across the Seattle area, California, New York and Texas.

Owens told GeekWire that Resect AI chose Washougal, across the Columbia River from Portland, for its community and access to technical talent. Resect AI plans to open a Seattle-area engineering and business office and aims to reach 50 employees by the end of 2026.

The unnamed investor group leaves an important part of the financing opaque. A lead investor can indicate whether a young AI vendor is being backed for technical research, an enterprise sales push or a later-stage consolidation strategy. Resect AI has disclosed the check without providing that market signal.

The capital still places Resect AI alongside a well-funded group of vendors trying to make AI systems measurable. Patronus AI announced a $50 million Series B on June 25th for evaluation, agent testing and simulation. Arize AI, Braintrust and WhyLabs sell adjacent observability, scoring and guardrail products.

Resect AI's proposed edge is its claim that intervention happens inside model operation instead of solely grading the output afterward. The next proof point is production evidence showing that this intervention works across models and workloads without degrading useful answers.

Regulation creates the opening

The timing gives Owens a ready enterprise sales argument. The European Commission began enforcing portions of the AI Act and new transparency requirements on August 2nd. The rules include disclosure and machine-readable marking requirements for certain AI systems and generated content.

Stanford's 2026 AI Index also found hallucination rates ranging from 22% to 94% across 26 models in one accuracy benchmark. The range reflects how sharply performance can change with the model and test design, which complicates any broad promise to eliminate hallucinations.

Interpretability and audit logs can support governance work, although installing a technical layer does not by itself establish compliance. Enterprises still need policies, human oversight, documentation and controls around where models are deployed. Resect AI is betting that those governance programs will require evidence from inside the model, giving Owens and his co-founders a place in the stack before AI systems reach customers and regulators.