Noeri raises $110 million to improve enterprise AI with human review
Edited and reviewed by AI Ping Editorial Team
3-line summary
- Noeri, founded by Stanford professors Yejin Choi and Carlos Guestrin, announced $110 million in funding.
- The company is building a system that uses work records to propose changes to enterprise AI and sends them to staff for approval.
- Menlo Ventures led the round, with NVIDIA among the investors. Noeri says its first partnerships are at the proof-of-concept stage.
Noeri, an AI startup co-founded by Stanford computer science professors Yejin Choi and Carlos Guestrin, announced $110 million in funding on November 16. Choi and Guestrin are co-CEOs. Their company is building a system that studies how a business uses its AI agents, proposes changes and sends them to a person for approval.
The founders say Noeri has begun proof-of-concept partnerships in business and scientific research. Companies can contact Noeri about improving an agent they already run or building a new one. The website has a form that asks what the company wants to work on. It gives no price list or date for general access.
Menlo Ventures led the financing. Madrona Venture Group, Kindred Ventures, Wing Venture Capital and NEA joined the round, with Cisco Investments, NVIDIA and Snowflake taking part as strategic investors. The $110 million is the amount raised across the round. The announcement did not give a valuation or amounts by investor.
A person reviews what the AI wants to change
Noeri says its system starts with the work an agent already does. Records of tasks, mistakes and corrections can show where an agent has trouble. A proposed fix might change the instructions or the tools the agent uses, or assign one step to a different model. Further training is another option.
The company says the system explains each proposed change and asks a person to approve it. It describes a process similar to code review. The agent can identify a change based on work it has done, and an employee can inspect the proposal before it alters the workflow in production.
Noeri says the system can work with both open-weight and closed models. It can revisit which model handles a step as new versions become available. That lets the company pursue improvements to one part of an existing workflow without replacing the whole system.
Costs and results measured on a company’s own work
Noeri says it measures quality and cost on the customer’s own traffic and against the measures the organization uses to judge that work. A team can compare changes on the jobs it actually sends to the agent. The launch materials give no customer-level figures for savings or gains in accuracy.
The company says customers retain ownership of their data and the workflows and models they develop, subject to the licenses of outside models. What a customer can move or run elsewhere depends on the third-party terms. The founders say they want what a company teaches its system to remain useful as it switches to a newer or cheaper model, including one from another provider.
Research that updates a model while it works on one problem
Stanford lists reasoning and new training methods for language models among Choi’s research interests. Guestrin co-founded Turi, a company that made tools for data scientists, and went on to lead machine learning and AI at Apple after it bought Turi. The two researchers have worked together on ways to train a model as it tackles a new problem.
Noeri points to a January paper co-authored by Choi and Guestrin. The researchers used the open-weight gpt-oss-120b model for tasks including mathematics and algorithm design. They continued training on results from attempts at an individual problem, so the model could use those results to change how it approached that problem.
In a separate paper co-authored by Guestrin, researchers studied how a model could learn from detailed feedback, such as a runtime error, alongside the outcome of a coding test. Both papers report results on specified research tasks. Neither provides a measurement of Noeri’s performance inside a customer’s business.
Noeri says it is working with partners from several industries, including some of its investors. It is based in Palo Alto, has an office in Zurich and is hiring researchers and engineers in both locations.
Sources
- November 16 launch and $110 million funding announcement — Noeri
- Product scope and company contact form — Noeri
- The Plural of Intelligence and the first proof-of-concept partnerships — Yejin Choi and Carlos Guestrin
- Research behind the company’s approach — Noeri
- Yejin Choi faculty profile and research — Stanford University
- Carlos Guestrin faculty profile and industry background — Stanford Engineering
- Learning to Discover at Test Time — Mert Yuksekgonul and co-authors
- Reinforcement Learning via Self-Distillation — Jonas Hübotter and co-authors