Resect AI takes $25M to fix hallucinations in-model
Resect AI launched with $25M to intercept LLM hallucinations at runtime instead of catching them after. How to evaluate the claim before you buy it.
LLM hallucinations just drew a $25 million bet on a different fix. Resect AI came out of stealth on September 3 with a claim worth taking seriously and checking carefully: instead of grading model output after the fact, its tooling watches the model's internals while it runs and intervenes before a fabrication reaches the user.
What actually happened
Per the company's launch announcement, Resect raised $25 million and is building both open-source and enterprise products to observe, detect, interpret, audit, and modify LLM behavior. Chief AI Officer Tim Walton's framing: technology that "observes exactly when and how models fail, and surgically fixes them." The backers were not named — the release says only that the money came from private equity investors.
The concrete part comes from SiliconANGLE. Resect built its own models first, then found the techniques carried over to open models including DeepSeek, Qwen, and Llama. It has published two Qwen3-based models on Hugging Face: a 600M-parameter fact-checker called Veritas and an 8B model. The 0.6B version scores 72.3% average on LLM-AggreFact, a 7.4-point improvement over base Qwen3. A GitHub repo is promised but not yet populated. The team is small — CEO Kevin Owens leads roughly four people out of Washougal, Washington.
Why hallucination tooling matters for your business
Read that benchmark again. 72.3% is a real gain over the base model and it is not a solved problem. Any vendor selling you "no more hallucinations" is selling a rate, not a guarantee — and the honest ones publish the rate.
That is the useful takeaway regardless of whether you ever buy from Resect. If your AI touches anything with liability attached — quotes, invoices, inventory counts, medical or legal text, anything a customer will act on — the question is not which vendor eliminates fabrication. It is what you do on the percentage that gets through.
Build that layer yourself and you own it. Ground the model in your own data instead of its memory, so a claim can be traced to a record. Make the model cite the row, order, or document it used, and fail the response when the citation does not resolve. Log every ungrounded answer and review the log weekly — that is your actual error rate, measured on your traffic, not on a benchmark.
A 600M fact-checker you can run yourself is a genuinely interesting building block for that. A promise that the problem is handled is not.
Key takeaways
- Resect AI launched September 3 with $25M to detect and correct hallucinations inside model internals at runtime
- Investors were not disclosed; the release names only "private equity investors"
- Two Qwen3-based open models are on Hugging Face — a 600M fact-checker (Veritas) and an 8B model; the GitHub repo is still empty
- The 0.6B model scores 72.3% on LLM-AggreFact, up 7.4 points over base Qwen3 — a real gain, not a solved problem
- Ground answers in your own records, require resolvable citations, and log ungrounded responses to measure your true error rate
No model is accurate enough to skip the verification layer. We build AI systems that answer from your data, cite the record behind every claim, and log what they could not ground — so you can see your real error rate instead of a vendor's. See how we build accountable AI, or tell us where a wrong answer would cost you money.
Sources: Resect AI press release, SiliconANGLE.
- #hallucinations
- #llm-evaluation
- #ai-reliability
- #funding
- #open-source
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
XDOF at $1.2B: the training data is the product
XDOF is in talks at a $1.2B valuation for collecting robot training data, three months out of stealth. Proprietary training data is the asset labs cannot self-serve.
Read itProofpoint's SOC analyst agent investigates, never acts
Proofpoint's SOC Analyst Agent runs security investigations on OpenAI Daybreak models but takes no action itself. The restraint is the design lesson.
Read it