The generative AI and large language model (LLM) explosion over the last few years has made AI agents popular. Although there is disagreement about the precise definition of artificial intelligence (AI) agents, most people believe that they are software applications with variable levels of autonomy that may be given tasks and make decisions.
To put it briefly, AI agents are more than just chatbots; they assist humans in completing tasks.
Even though it’s early, companies like Google and Salesforce are already making significant investments in AI agents. Andy Jassy, the CEO of Amazon, recently alluded to the possibility of a future Alexa that is more “agentic,” meaning that it will be more action-oriented than spoken.
Startups are also profiting from the hoopla at the same time. The most recent of these is Juna.ai, a German business that aims to “maximize production throughput, increase energy efficiency, and reduce overall emissions” by automating intricate industrial operations in factories.
In order to accomplish that, the Berlin-based startup said today that it has raised $7.5 million in a seed round from Sweden-based Norrsken VC, Silicon Valley venture capital firm Kleiner Perkins, and Kleiner Perkins chairman John Doerr.
The method is self-learning
Christian Hardenberg (shown above, right) and Matthias Auf der Mauer (pictured above, left) founded Juna.ai in 2023. Hardernberg was the former chief technology officer at the massive European food delivery company Delivery Hero, and Der Mauer had previously developed a predictive machine maintenance startup called AiSight, which he sold to the Swiss smart sensor company Sensirion in 2021.
Juna.ai’s primary goal is to assist manufacturing facilities in becoming smarter, self-learning systems that can increase profits and, eventually, reduce their carbon footprint. The business specializes on so-called “heavy industries,” which include sectors like steel, cement, paper, chemicals, wood, and textiles that utilize a lot of raw materials and have extensive production processes.
Der Mauer told TechCrunch, “We work with very process-driven industries, and it mostly involves use-cases that use a lot of energy.” “So, for instance, chemical reactors that generate something by using a lot of heat.”
The software from Juna.ai examines all of its past data collected from machine sensors and interfaces with the production tools used by firms, such as industrial software from Aveva or SAP. This could include temperature, pressure, velocity, and all of the output’s measurements, including thickness, color, and quality.
By providing operators with real-time data and direction to guarantee everything is operating at maximum efficiency with the least amount of waste, Juna.ai assists businesses in training their internal agents to choose the best settings for machinery.
For instance, a reactor may be used to combine various oils and subject them to an energy-intensive combustion process in a chemical plant that generates a particular type of carbon. Conditions, such as the amounts of gases and oils used and the process temperature, must be ideal in order to maximize production and reduce leftover waste. According to the theory, Juna.ai’s agents advise the operator on what adjustments to do in order to get the best results by using historical data to determine the optimal settings and accounting for current conditions.
Businesses can increase throughput and lower energy consumption if Juna.ai can assist them in optimizing their production machinery. It benefits the customer’s bottom line as well as its carbon footprint.
According to Juna.ai, it has developed its own proprietary AI models with open-source programs like PyTorch and TensorFlow. Additionally, Juna.ai uses reinforcement learning, a branch of machine learning (ML), to train its models. This method involves a model learning by interacting with its surroundings; it takes various actions, observes the results, and becomes better.
Hardenberg told TechCrunch, “The intriguing thing about reinforcement learning is that it’s something that can take actions.” Conventional models merely make predictions or sometimes produce anything. However, they are powerless.
A large portion of Juna.ai’s current functionality is more like that of a “copilot”; it displays a screen that instructs the operator on what adjustments to make to the controls. However, allowing a system to perform real actions is useful because many industrial procedures are quite repetitious. To guarantee a machine maintains the proper temperature, for example, a cooling system may need to be adjusted on a regular basis.
PID and MPC controllers are already widely used in factories to automate system controls, thus Juna.ai may potentially accomplish the same. However, selling a copilot is easier for a young AI business; for the time being, it’s baby steps.
Technically, we could let it to operate on its own at this time; all we would have to do is establish the link. But ultimately, it all comes down to gaining the customer’s trust,” der Mauer stated.
Hardenberg went on to say that manufacturers are already “quite efficient” at automating manual operations, therefore the startup’s platform isn’t particularly useful for labor savings. It ultimately comes down to streamlining those procedures to reduce expensive waste.
“Compared to a process that costs you $20 million in energy, there isn’t much to gain by removing one person,” he stated. “The true benefit, then, is whether we can increase energy spending from $20 million to $18 million or $17 million.”
Agents with prior training
The main promise of Juna.ai at this time is an AI agent that is customized for every client based on their past data. However, in the future, the business intends to provide “pre-trained” agents that are off-the-shelf and require little training on the data of a new customer.
“If we repeatedly construct simulations, we eventually reach a point where we may have reusable simulation templates,” der Mauer stated.
For example, it may be feasible to move and elevate AI bots between clients if two businesses employ the same type of chemical reactor. The general idea is one model for one machine.
It is impossible to overlook, though, that businesses have been reluctant to fully embrace the emerging AI revolution because of worries over data protection. Juna.ai is unaware of these worries, but according to Hardenberg, it hasn’t been a significant problem thus far, in part because of its data residency limits and in part because of the assurance it provides users that latent value may be uncovered from enormous data banks.
“We leave all data in Germany for our German customers, so it hasn’t been a big problem yet, but I was seeing that as a potential problem,” Hardenberg stated. “We provide excellent security guarantees, and they set up their own server. Although they had a ton of data laying around, they haven’t been very successful in turning it into something useful; alerts and possibly some manual analytics were the main uses. However, we believe that we can do much more with this data—create an intelligent factory and use the data to become its brain.
A little over a year after its founding, Juna.ai already has a small number of clients, though der Mauer stated he is not yet in a position to disclose any names. However, they are all headquartered in Germany and either have operations abroad or are subsidiaries of foreign-based businesses.
Hardenberg continued, “We intend to grow with them — it’s a very good way to expand with your customers.”
Juna.ai is now well-positioned to grow beyond its current staff of six, with intentions to double down on its technical competence, thanks to the new $7.5 million in the bank.
“In the end, it’s a software company, which essentially means people,” Hardenberg stated.