AI agents are all the rage, a trend fuelled by the generative AI and large language model (LLM) boom these past couple of years. Getting people to agree on what exactly AI agents are is difficult, but most contend they are software programs that can be assigned tasks and given decisions to make - varying degrees of autonomy aside.

Simply put, AI agents do more than a simple chatbot: they help humans accomplish things.

It's very early days still, but well-established players like Salesforce and Google are already making very heavy investments in AI agents. Speaking of their own Alexa, Amazon CEO Andy Jassy recently dropped a hint on a more "agentic" version of the voice assistant in the future—more action and less words.

In tandem, venture capital funds are also lining up to invest off the hype. The latest of these is German company Juna.ai, which wants to help factories be more efficient by automating complex industrial processes to "maximize production throughput, increase energy efficiency and reduce overall emissions."

And to get there, the Berlin-based startup said today that it has raised $7.5 million in a seed round from Silicon Valley venture capital firm Kleiner Perkins, Sweden-based Norrsken VC, and Kleiner Perkins' chairman John Doerr.
The way of self-learning
Founded in 2023, Juna.ai is the brainchild of Matthias Auf der Mauer (above left) and Christian Hardenberg (above right). Der Mauer founded an earlier predictive machine maintenance startup called AiSight, which he then sold to Swiss smart sensor company Sensirion in 2021, while Hardernberg previously served as chief technology officer at European food delivery behemoth Delivery Hero.

At its core, Juna.ai is hoping to help manufacturing facilities become smarter, self-learning systems offering better margins and, ultimately, a reduced carbon footprint. So, the company hones in on "heavy industries" — industries such as steel, cement, paper, chemicals, wood, and textile, where production processes are undertaken on a large scale, consuming tremendous amounts of raw materials.

“We work with very process-driven industries, and it mostly involves use-cases that use a lot of energy,” der Mauer told TechCrunch. “So, for example, chemical reactors that use a lot of heat in order to produce something.”

Juna.ai’s software integrates with manufacturers’ production tools, like industrial software from Aveva or SAP, and looks at all its historical data garnered from machine sensors. This might involve temperate, pressure, velocity, and all the measurements of the given output, such as quality, thickness and color.

With this understanding, Juna.ai empowers businesses to train their in-house agents so that they can determine the best settings for the machinery, giving operators real-time data and guidance so everything will run to peak efficiency with little or no waste.

As a simple example, a chemical plant that produces a very coveted form of carbon may operate a reactor that it feeds a variety of oils together with and applies an energy-intensive combustion process to. To attain optimal levels for maximum output and minimum residual waste, conditions such as gas and oil levels and temperature on application should be at the optimum. With the use of historical data that has set optimum settings and knowing the real-time conditions of the process, the agents of Juna.ai are said to advise the operator on what adjustments to make in order to get the highest output.

If Juna.ai can fine-tune production equipment, then companies can increase throughput without taking up more energy. The bottom line for the customer and the carbon footprint both benefit from it.

Juna.ai claims to have developed its proprietary AI models using open-source tools like TensorFlow and PyTorch. To train their models, the organization is using reinforcement learning-a subset of machine learning wherein a model learns by interacting with its environment. It tries different actions, observes what happens, and improves upon those findings.

The interesting thing about reinforcement learning is that it's something that can take actions," Hardenberg said to TechCrunch. "Typical models only do predictions or maybe generate something. But they can't control.".

Much of what Juna.ai is doing today has more of a "copilot" character: it lays out a screen telling the operator what adjustments they should be making to the controls. However, many industrial processes are just flat-out repetitive, so the ability to enable a system to take real action is useful. A cooling system, for example, must constantly be adjusted in fine degrees to maintain a machine at the correct temperature.

This is something factories are already very accustomed to doing with PID and MPC controllers. The AI startup could very well do this, but for a young AI, it is easier to sell a copilot-behavior-that is baby steps for now.

"It's technically possible for us to let it run autonomously right now; we would just need to implement the connection. But at the end of the day, it's really all about building trust with the customer," der Mauer said.

According to Hardenberg, the advantage of the startup's platform doesn't relate to labor savings; factories are pretty efficient already in trying to automate those manual processes. It is rather about optimization of those processes in order to cut down on costly waste.

"There's not a lot to gain by removing one person, compared to a process that costs you $20 million in energy," he said. "So the real gain is, can we go from $20 million in energy to $18 million or $17 million?"
Pre-trained agents
For now, the big promise of Juna.ai is an AI agent tailored to each customer, built from their historical data. The company, however, plans to offer, in the future, off-the-shelf "pre-trained" agents that require little training on new customer data.

"If we build simulations again and again, we get to a place where we can potentially have simulation templates that can be reused," der Mauer said.

So if two companies have the same type of chemical reactor, for example, then possibly lift-and-shift AI agents between customers. One model for one machine is generally the case.

However, one can hardly deny the fact that enterprises have been shy of jumping headlong into this bursting AI revolution because of concerns for data privacy. A concern lost on Juna.ai, but partly because of its data residency controls, and partly because of the promise it gives customers in terms of unlocking latent value from vast banks of data, Hardenberg does not seem to worry about this as much of an issue so far.

I was viewing that as a potential problem, though so far it hasn't been such a big problem because we leave all data in Germany for our German customers,” said Hardenberg. "They get their own server set up, and we have top-notch security guarantees.". From their side, they have all this data lying around, but haven't been as effective in creating value from it; it was used mainly for alerting, or perhaps a bit of manual analytics. But our view is that we can do so much more with this data-build an intelligent factory, and become the brain of that factory based on the data that they have.

A little more than a year since its foundation, Juna.ai has a handful of customers already, though der Mauer said he’s not at liberty to reveal any specific names yet. They are all based in Germany, though, and they all either have subsidiaries elsewhere, or are subsidiaries of companies based elsewhere.

“We’re planning to grow with them — it’s a very good way to expand with your customers,” Hardenberg added.

With the new $7.5 million in the bank, Juna.ai is now well-financed to expand past its current six-headcount and double-down on its technical expertise, according to an interview with the company's founder, Christopher Hardenberg.

"It's a software company at the end of the day, and that basically means people," said Hardenberg.