Josh Knutson and Ryan Thill are building Lium for a problem that sits just outside the usual AI demo.

Most AI tools are very good at text, code, and spreadsheets. Lium is focused on the messier stuff, the huge physical-world data sets that sit inside farms, climate labs, energy systems, logistics networks, and other operations where the answers are buried under terabytes of data.

Knutson, the CEO and co-founder, describes Lium as an “agent harness” or a cloud operating system for agents. The idea is to give language models the tools they need to work over large, complex data sets that they cannot handle well out of the box. Instead of asking a data scientist to build a pipeline every time someone has a question, Lium lets subject matter experts ask questions in natural language and then builds the tools and workflows needed to answer them.

Thill, co-founder and president, said the core user is often not a software engineer. It is the person who knows the domain, knows the data matters, and knows there are answers inside it, but cannot easily get them out. He gave the example of a farm operator working with soil reports, NOAA data, tractor data, and crop performance information. The operator may know something is off, but does not have the time or technical skill to combine all those sources into a useful answer.

That is where Lium is meant to fit. A user can describe what they want to know, and the system builds repeatable workflows around the data. Once those workflows exist, other people inside the organization can use them too. An analyst can build the tool, and a CEO can later ask a simple question that relies on the analyst’s work in the background.

That shared layer is one of the more interesting parts of the product. Knutson described work with the North Carolina Institute for Climate Studies, where scientists and researchers built tools inside Lium, then on-screen meteorologists could ask questions and get answers using the right climate data without needing to understand every data source underneath.

The company’s bet is not that AI replaces the expert. It is that AI needs the expert. Knutson said Lium is built around human-in-the-loop workflows because language models do not have enough training data to understand all the hidden patterns and details inside many physical-world data sets. The system has to know when to stop, ask the human for domain knowledge, and then turn that knowledge into a tool the system can use again.

That point matters because the obvious fear is job loss. If a person’s job is to build reports, what happens when anyone can ask Lium for the same report? Knutson and Thill argue that the expert becomes more valuable, not less, because the tool captures and scales their knowledge. Thill compared it to software engineers using AI coding tools. The tools make people more productive, and that can create demand for more work that was not worth doing before.

Lium is still early, but the founders say they are seeing strong interest. During private beta, around 50 groups worked in the platform. Now that it is public, the challenge is different. Instead of onboarding users by hand, the company has to explain the product clearly enough that people can find it, understand it, and get value without a sales call.

That is not easy, because Knutson and Thill say many potential users do not know this kind of system is possible. For Lium, the main competitor is not another startup. It is the belief that this kind of data is too hard to work with.

Fundraising followed a similar path. Knutson said early investors were skeptical because he and Thill did not have the usual Silicon Valley AI profile. They had startup experience, but not the standard AI pedigree. The company raised a smaller pre-seed round than it wanted, then came back after showing it could build things people did not think it could build. That proof changed the conversation.

The company spent roughly 18 months learning and building before going public. Knutson said this was not the kind of product where you can ship a tiny version and see what happens. If someone brings a terabyte of data, the system has to work. That meant building alongside design partners until the product was strong enough to handle real use.

Now the work is public. Lium is learning from users, tightening the funnel, and building around what people actually do with the product. The name, by the way, comes from language plus the suffix of physical elements, a nod to the company’s goal of connecting language to the physical world.