Edge AI That Cuts Chemical Waste

What if a lab test that takes 12 to 24 hours could be replaced by a live estimate that guides dosing in real time? We walk through a high-stakes water story where boron control in desalination demanded more than a clever model—it needed a secure, local-first AI system that works across wildly different plants.

Our journey with Acciona started with a simple idea: a virtual sensor to predict boron and avoid overusing caustic soda or risking fines. The reality was messy. Membranes, sensors, and SCADA setups varied from site to site. Cybersecurity kept data locked on-prem, and lab workflows produced sparse, noisy labels. A single global model wasn’t resilient enough. So we flipped the playbook and orchestrated many models at the edge—one per rack when needed—packaged in Docker, deployed with a click, and monitored locally with InfluxDB and Grafana.

We break down the full stack: MQTT brokers to standardize data, connectors for heterogeneous OT systems, TensorFlow for inference, and JupyterLab plus MLflow for on-device training and versioning. This architecture kept raw data inside the plant while a cloud console managed applications securely. The payoff was immediate: accurate boron estimates tightened dosing, cut chemical spend, reduced penalties, and built operator trust by showing predictions alongside lab results. One site saved over $200,000 in a year; scaled across the fleet, the impact reaches well into the millions, with healthier water as a bonus.

Beyond boron, the same edge AI approach unlocks energy optimization for high-pressure pumps, membrane fouling detection, and even computer vision tasks—without compromising critical infrastructure security. If you care about industrial AI that actually ships, this is a practical blueprint: local models, secure orchestration, and a path from pilot to fleet.

Enjoy the episode, share it with a teammate who wrestles with on-prem constraints, and leave a review. Want to see it live? Ask us for the free trial and we’ll set up a demo.

Transcript

Setting The Stage And Company Origin

SPEAKER_00 0:06

Hello everybody. Okay, so it has been a long day, right? So today's gonna be a little more like a story on a real use case that we did like three years ago, even with before the HAI existed as such, right? So it started with uh yeah, here it is, with a company. But before get into our is that who is Barbara? Now probably you will be all wondering what is this company with this funny name, so it's not somebody's not the founder.

Why “Barbara” And Cybersecurity Roots

SPEAKER_00 0:40

It's uh long story short, is because we started a company based on the cybersecurity. And there's an old saying on in Spain that you only remember St. Barbara when it's an issue, right? So you only remember cybersecurity when there's uh when there's a problem. And and after that, so we've been evolving the company because of the customer request, and we have evolved it into what is HAI.

Introducing Edge AI Platform Capabilities

SPEAKER_00 1:07

And what we do now is that we are able to deploy the uh we are engaged management platform that allows to deploy AI models across many different uh thousand or hundreds of different devices in a very cyber secure way, and and also includes a lot of uh infrastructure around it for the industrial environment. So you have an ecosystem of applications that allow to create these use cases that are important.

The Client: Acciona And Water Focus

SPEAKER_00 1:37

So as I was saying, it all started with uh uh this use case about HAI and journey into that with a customer, it's a big customer in Spain called uh Actiona. You may not know it's uh it's a company that covers energy, construction, and other things, and also water. So they created in 2019 uh uh digitalization uh department whose idea was to bring innovation into the water,

Desalination Process And Boron Problem

SPEAKER_00 2:06

right? So they cover the whole water cycle, but in this case, uh and uh but they have an issue in a particular use case because the desalination plants. So the salination plant is a funny thing. It looks like uh, okay, the water is everywhere, so it should be quite easy. You just remove the salt, and that's it. And basically that's what we do. So they get the water, first they pre-treat it, they bring, they take all the organics out. They go through a process which is called reverse osmosis, and they take the all the salt out, but there's still other products in the other chemical products into the uh desalination process. And what they do is they treat it afterwards with some chemicals, right? Uh one is difficult to get rid of is boron. Okay. So the boron, okay, say you there's a way of treating it, you just put uh caustic soda on it and then you treat it. But you know you need to know how much uh boron is on the water, otherwise you may have issues. It may be bad for the health, or even you can get penalties.

Lab Delays And Need For Virtual Sensor

SPEAKER_00 3:15

So they say, okay, let's let's analyze it. But when they analyze it on the lab, it takes normally 12 hours, 24 hours in order to know exactly what is the amount. And when that happens, they cannot treat it because that's the water is already gone. So they have to find a way of uh of improving it, right? So they say, okay, how can we estimate the amount of boron that's in the water so we can treat it more precisely? And even better, can we predict how much boron are we gonna use? And then there's another thing that they also thought about doing is energy optimization because there's a huge amount of energy they need to put into getting that. But that's gonna be the second stage. So originally say, okay, they had this idea, no, they had this vision. Say we're gonna create a virtual sensor for boron. That's quite simple, right? It doesn't look like a very complex model anyway. And but still, they have to do it, and they say, okay, they get all the things that are already, you know, the pH, uh turbidity,

Data Quality, Variability, And Security Hurdles

SPEAKER_00 4:18

or temperature, all that. That's how they get all the data, they got the results from the lab. And say we should be quite straightforward. But they also thought, okay, we're gonna do it, we're gonna deploy to all the 1989 plants that we have in the world. But there's an issue here because the the plants need to be, you know, it's a critical infrastructure, so the data cannot live there. And they cannot leave the premises, they cannot leave the plan. Okay, so well, let's see how we're gonna do it. So they they say, okay, we'll figure that out later, but that's our vision, that's what we want to achieve. So they started and they found other challenges as well. That even that they had the data, it's difficult to get. Because in the lab, one of the main issues is that, okay, they have the they they maybe do uh one measure per day or twice. So, in order to have enough data, you imagine how much years you need to gather the data out of that. Also, the quality of the data, no, because there goes the guy to the lab, get the sample, they put it in the lab, they write in an Excel sheet the time. So that was it was done until so it was also an issue. They have to clean a lot of data before that. Also, a problem with cybersecurity. There are plants where you cannot connect, you cannot get the data outside of the plant. You actually to get all the data, they have to ship the data in a hard disk in a special place to the laboratory to do everything. Another issue is that there's a lot of variability because of the business. They don't have the same plant everywhere. It's difficult. It is different in each one of them because they have concessions, one have been built by somebody else, not from them. One has some system, others have other ones. So that's

First Model Built And Deployment Blockers

SPEAKER_00 6:08

difficult. So they were all the challenges that they say, but they say, okay, we'll figure that later. First of all, let's try to build the model. So that's what they did. Okay, that's gonna be quite straightforward. So they went to the lab and they just installed a computer and they put a mod, they create the model, they had to clean up the data and all that. It took them some quite a long time, but they managed to do it. They clean it, they they create the model, they say, okay, now I need to connect to the SCADA system. So, okay, I need to find a way of connecting it. So uh it was a pain to say. Well, but they finished, they did it, okay. Now let's deploy. We have this model, it's working here. Actually, we tested, it works very well the model. Let's deploy it somewhere else. And and that was impossible. They said we need to find to find a way to do that.

Bringing In Barbara For Secure Orchestration

SPEAKER_00 6:58

So they called us. No, they called Barbara. And uh the advantage is that we they knew us for some reasons, so we managed to sort out different types of problems that they had to deploy these models. These were not the problems or um creating the model itself originally or having the model. It were other types of issues, no? And here what we sort for them is that we have a central platform that allows them to allow you to get to manage all the different equipment that they have across the different plants. So they sort already something, it's gonna be easy to deploy. We are based on Dockers, so that's all another advantage because they say, okay, it's gonna be quite standard to deploy everywhere, so I don't need to take care of everything. We can uh it's a secure orchestration. We take care of the operating system, we take care of the cybersecurity. So we also take care of uh upgrading of the model and all that. We have all the connectors to connect to the different systems, and for them already also sort that part out because it's getting the real data is complicated. And then we at the time we didn't have it, but now we do. But basically, it was uh we were able to deploy the mall ops there. So, more or less, this is our platform

Edge Architecture And On-Prem Data Control

SPEAKER_00 8:15

here. You see, for a central panel, you see all the different devices that are installed across the wall. You can go inside of them and you can deploy the application in a single click. Okay, you can deploy a model in a single click. And also we have a marketplace full of applications that allows you to do that. All the all the uh necessity connections, everything that is around, getting the data, exploited it, and in an industrial environment. Because the industrial environments are quite specific. So they they have their complications. So we have that part, so then we sort it. And what we do? So the way we propose the solution, it's something like something like this. Okay. We as I said before, we are just an orchestration platform. So we don't actually see the data. That's important for cybersecurity. All the data stays on the plans. The only thing that is on the cloud is the orchestration. So we only see that there are things and there are applications that are installed. So you hear everything here is installed in an edge device. These are bigger edge devices than all these microcontrollers you're talking about. So uh some of them were from AdvanTech, another from from Westnater. So we we create that

Centralized vs Local Training Tradeoffs

SPEAKER_00 9:35

part. So they put this device here, and then you then can deploy the models, or your own Docker applications, you can deploy the straight away the models, and a lot of other applications that you can deploy, even someone from uh also IE, like YOLO, YAMA, all that, you can deploy stata weight and single click to all the devices. And thanks to this type of architecture, and because of the flexibility of the platform, allows them to create different types of use cases. So they say, okay, how we're gonna do it? They have three different types of models to deploy. Originally it was thought to do like uh uh you know a centralized training, so because they already have the model centralized, say we're gonna deploy all the plants. That was the first idea. But then they realized that they have an issue. Every single plan was different. So the membrane is different, the sensors are different. So the model that they have in the cloud doesn't work. They need to have a different model on every plant and a different model for every rack. Kind of different work. Okay, so one idea was okay, let's do it centralized,

Standard Edge Stack For Industrial AI

SPEAKER_00 10:42

a different model, then they have to create 180 times 10, so 900 different models in the cloud. That was a pain. So, okay, we can do something different. Why couldn't we do it local? And that's what we propose. So we propose a solution that allows them to deploy locally the and to train the models locally. I get there. And also uh there's another option of being federated, but although we haven't done it, but basically it's similar, no? You train the uh centralized, you send it uh to the on-premises, do the inference, then you send the data back to the to the cloud, and you retrain the model that you send over there. But it was not the case for them, it was more a local training. So, in order to do that, um we do this type of architecture that is quite standard. This is not only for them, I will go to them later. But basically, it's here. Normally we have always this type of architectures that make it very flexible to deploy the models on there. So we put a local MQTD broker, this is an H node, right? Everything is in the plan, one of one of the racks. You put the MQTT broker there, from there you create the connectors. Here is one, but you can have any type of connectors. So all the data is standard there. And then from there you connect

Local Training Per Rack And Monitoring

SPEAKER_00 12:02

the model. So the advantage is you have a single place where all the data is concentrated, and from here you can send it also then to another database, you can send it to the cloud. This is not for them, but this is more like a generic. You have an ML monitoring tool that we also see, so you can see whether you are discontinuing. So this is applications that normally at Barvara, the way we deploy normally these models or how we um make them available for everybody, is this type of architecture. And that's very flexible because they allow to deploy different models in the same device. So um let me go. So, what did we do for them? There was a little swap here. So uh finally, what they started is that the sensor, the debroker, tensor flow, we have uh an influx database for time series locally, and a grafana, it's also to see everything that's happening on the model. And then for the local training, that's the one that we're talking before. We put also a Jupyter Lab here uh notepad with a Python model that they had it, an ML flow that allows them to have different versions of the model, then they can compare it with the real data at the end of the day, with the and then they can switch and select the best one. So we are doing for them a local train uh training uh locally for every one of the racks on every one of the plants, and that allows them to create a specific model for them. So that's another interesting use case is for HAI. It's not only reducing data um to the cloud, it's not only uh uh lowering costs, it's because you can do it specifically at a different model per each device because the environment is different everywhere.

Outcomes: Savings, Trust, And Scale

SPEAKER_00 13:53

So what they manage is that what I mentioned. They put the model, they have uh now they are able to do automation uh automated estimation of boron. Now, thanks to that, they are reducing the amount of chemicals they need to use. And also another advantage is that they can manage all these devices in all the plants locally. And now they have like um this is a little bit old now, but they they see more or less all everything here, the different things that are um uh reading, and then they can compare with the real one, what is predicted, what is real, so they they can be sure. And also it gives the all the people in the lab, in the lab in the plant the confidence that this AA thing that for them is quite science fiction is real because okay, actually it's actually it's happening, it's it's doing what it was supposed to do. So that's uh really helps. So how was it the end? The impact. So yes, super planning one year, they saved more than $200,000 in reducing of chemicals and fines. That's important. And also the health improves. So in three years, it's more than if you put it everywhere, it's like more than 70 million dollars of savings that you may have. And also help them to go much, much faster to deploy everywhere. So, well, basically that was the idea.

Why Barbara: Security, OT Fit, And Trial

SPEAKER_00 15:22

So now because I'm uh SAT and the VP of sales, so I need to do some sales selling here. So why Barbara? Actually, we are a great solution because allows every uh the industrial companies to have security that the they're gonna we're gonna we're gonna have distributed equipment that is secure and we can manage from a central platform, and we are very uh focused in the cybersecurity. That's this thing I6 ASC62443 uh recommendation of cybersecurity that makes that uh uh the the equipment has everything is encrypted, encrypted boot, no open ports, all that. So we comply with that part. That's the part with Barbara that we mentioned before. We are edge native solutions, so we are cloud agnostics, you can integrate with any type of cloud. We are hardware agnostic, we can deploy any type of hardware, normally any type of micro uh microprocessor hardware, no microcontrollers yet. We are tailored for the OT environment, so we have more than 100 OT applications ready. So we can be very flexible, we can integrate with different systems, we can deploy other tools. I didn't mention, but that was the first use case, but now they were putting tools on top of that. They are using our infrastructure also to deploy the traditional OT tools that they do, and for them it was a pain to do that everywhere. So now from a central point, they can put it everywhere. We are AA focused, actually. We we were doing this like three years ago. We were HAI before anybody was like, it was a buzzword at the moment. We'll say, okay, we are doing already AI at the edge. And it was interesting because it's it was driven by the customer needs. It was not something that we thought it was a good idea. And also UX orientation. Um it's a pity I haven't been able to do a demo because well, the computer is over there and they don't allow me to do it. But it will be great if you guys talk to me and we go and do a demo. Even if it's an OT tool, it's very easy to deploy. Deploy an application, deploy A models, all that. It's a single clip and it's already embedded. So, well, still we have a free trial here. So if anybody wants to access that and connect, we will be gladly um do uh a demo for you, create a trial account so you will be able to access it. So that's it. Thank you.