Lium, a startup formerly known as Astromind, today announced the launch of an “agentic harness” that helps large language models dig into the most complex and messiest datasets.

The launch comes after it closed on $5.5 million in seed funding from investors that included SJF Ventures, Wavemaker 360, Reach Capital and GC&H Investments.

The Dallas-based company is tackling a problem that continues to befuddle even the most advanced artificial intelligence models. Though LLMs excel when it comes to parsing text and code, they struggle immensely with more complex scientific datasets. For instance, things such as satellite imagery, seismic surveys and electromagnetic spectrum analysis confound even the most advanced reasoning models.

There are workarounds, but these methods are extremely tedious, forcing scientists and engineers to become ad hoc developers and manually prepare their data so it can be analyzed.

It’s really a question of data legibility, and it’s one that Lium intends to solve by ingesting the most challenging data and spitting it back out into a format that LLMs can understand. Its platform acts as a kind of connective tissue between complex datasets and the reasoning capabilities of LLMs. It automatically restructures the information in such a way that it can quickly be parsed through, enabling scientists to ask and receive answers informed by raw telemetry.

Co-founder and Chief Executive Josh Knutson said the limitations of LLMs are most apparent when it comes to understanding the physical world. “AI holds huge potential to solve many of humanity’s most pressing problems, but the most important data across energy, science and infrastructure remains difficult to reason over,” he explained. “Lium helps teams work with their own data to get better answers, faster, and make that a permanent capability. We’ve created the agentic harness purpose built for turning complex data into actual knowledge.”

Lium automates the translation of these hard-to-understand datasets using AI agents. For each new data type it comes across, it creates a custom agent that’s able to understand it. These agents then get to work by structuring the raw information into a format that LLMs can ingest, so that when someone queries the model, it will deliver consistent and reproducible results.

If true, this is a big deal, for hallucinations are incredibly common when LLMs attempt to dig into complex numbers they cannot easily navigate. What’s more, Lium says, its agents get better over time. As chatbots query the revamped dataset, the agent learns to refine its formatting techniques so that the data becomes more searchable.

Knutson said Lium’s technology has already shown good results in high-stakes environments. The startup has been working with astrophysicists to dig up insights from notoriously byzantine sparse X-ray data. In one case it analyzed the atmospheric makeup of exoplanets so well that it could assist with the search for extraterrestrial life.

The company is also applying itself to Earthbound tasks such as processing terabytes of satellite images and weather station data for the North Carolina Institute for Climate Studies. Meanwhile, the industrial power generator services company nexGEN Inc. tapped Lium to automate electromagnetic spectrum analysis, turning raw data into consistent generator health reports and replace a tedious, manual process.

Lium’s other co-founder, Ryan Thill, said much of the world’s most important data has proven to be virtually inaccessible to LLMs. By making it legible, it can unlock many more exciting use cases for AI.

“Lium creates a centralized interface where anyone can get the results they need — whether they’re a scientist wanting to more thoroughly examine soil conditions or a CEO who needs to build a financial model directly from those findings,” he said. “We’ve already seen the possibilities this accessibility can unlock through our work with astrophysicists. In one instance, our platform enabled the precise atmospheric analysis that could one day help identify life beyond Earth.”

Image: SiliconANGLE/Gemini