Resect AI Raises $25M to Try Changing Enterprise Models Before Bad Answers Reach Users

Resect AI has emerged from stealth with $25 million to build tools it says can spot hallucinations while a language model is running, then alter the model’s behavior before an unreliable answer reaches an enterprise user. The promise reaches beyond conventional after-the-fact answer checks—and remains to be proven in deployed workloads.

The financing came from unnamed private-equity investors. Resect said it will use the money for research and development, go-to-market work, and hiring in the Seattle and Portland markets. The company did not disclose a financing stage or valuation, leaving both the structure of the round and its backers opaque.

A bet on changes before the answer

Resect calls its intended product an accountability layer for large language models. The company says it is developing technology to observe, detect, interpret and audit model behavior, as well as modify that behavior to reduce hallucinations—false or fabricated responses presented as answers.

Its proposed mechanism is intervention during inference, the process of producing a response. Resect says the tools are meant to inspect model activity, identify emerging hallucinations, modify behavior and retain audit records. That is a larger claim than detecting a bad output after generation: it requires deciding quickly which behavior is failing and changing it without derailing an accurate answer.

What Veritas can already be inspected for

The public Veritas releases are 0.6-billion- and 8-billion-parameter fact-checking models based on Qwen3 and licensed under Apache 2.0. The 8B model repository describes the model as fine-tuned for fact-checking and factual-consistency verification.

Resect reported that its 0.6B Veritas model reached 72.30% average balanced accuracy on LLM-AggreFact, compared with 64.93% for the underlying Qwen3 model in non-thinking mode. The company said the test set was unseen during training. Its 8B model card reports 75.47%, versus 73.17% for Qwen3-8B in the same mode.

LLM-AggreFact aggregates 11 human-annotated datasets on fact-checking and grounding, according to the 8B model card. Those figures are useful, narrow evidence of the company’s factuality work, not a demonstration that its enterprise product can alter a running model across customer systems. The card also notes that performance can vary with hardware configuration and the vLLM version.

vllm serve "resect-ai/veritas-8B-fact-checker-non-thinking-1.0"

The proof obligation moves to production

NeuroWave is intended as the enterprise audit product built on Resect’s internal-visibility tools. The planned feature set spans detection, behavioral modification, audit logs, vLLM integration and agent controls. For buyers in regulated or accuracy-sensitive work, that combination could make the product a governance layer as well as a reliability tool—but those capabilities are company descriptions while the suite remains gated behind a waitlist.

The funding gives Resect room to pursue that proof. But the central question is not whether a specialized model can improve a factuality benchmark. It is whether the company can show that real-time intervention works across enterprise models and workloads while preserving correct, useful responses. That distinction separates an inspectable open model release from a production accountability product.