From PEDC to APEX: Banyan.eco’s Disruptive AI for EMS and OEM Companies

Estimated reading time: 9 minutes

The electronics industry’s struggle with supply chain resilience and increasing regulatory complexity has created a costly compliance challenge for OEMs and EMS providers alike. At the same time, advances in artificial intelligence are opening new possibilities for automating some of the industry’s most data-intensive tasks. At the Pan-European Design Conference (PEDC), Banyan.eco cofounder Francis D’Souza introduced a new approach that combines AI-driven data acquisition with a deterministic verification layer to eliminate compliance risk from AI hallucination and dramatically improve productivity.

In this interview, Francis discusses Banyan.eco’s origins during the COVID-era supply chain crisis, the growing burden of substance compliance, and the results of a white paper study he and his co-founder will present at APEX EXPO showing how their technology will transform compliance workflows and productivity.

Marcy LaRont: Francis, I've been tracking Banyan.eco for a little while now. Please briefly introduce yourself and the company.

Francis D'Souza: My name is Francis D’Souza. I am of Indian origin, I have a Portuguese-sounding name, and I live in Paris. I have had my career in India, Europe, and the U.S. with companies like Siemens, Thales, and Ericsson in various roles across IoT, semiconductors, and product, sales, marketing, and strategy.

LaRont: You're the quintessential international citizen. When and how did Banyan.eco get started?

D'Souza: During COVID, we figured out that even the small capacitor that costs a fraction of a cent can bring a massive production line to a halt. Later, in mid-2023, we began to see breakthroughs in AI technology and a glimpse of what it could solve.

Since that time, the electronics industry has struggled to become more resilient amid complex supply chains. Having lived through the COVID challenges just a few years before, my cofounders—Zakaria (Zak) Bouachra, CTO, and Alizée Dubois, Chief Customer Officer and I began examining how we could leverage the power of the LLM breakthroughs. We wanted to make it applicable and practical for the electronics industry, so that the industry spends less time chasing, gathering, and maintaining data, and more time on value-added activities. Alizée had been the sustainability manager at an electronics company and realized how difficult it was to get the data needed to make BOM changes and verify compliance. Zak was on the machine learning and data side. We wanted to determine how to put it all together, leverage the technology, and make a difference in the industry. That's how Banyan.eco got started.

LaRont: I'm still fascinated with the “COVID effect,” both the problems and solutions that came directly out of it. Tell me about your presentation at PEDC.

D'Souza: Zak and I gave a presentation, “Agentic AI to Drive Down EMS and OEM Compliance Costs.” An electronic design includes lots of components that undergo various industrial processes and are part of a complex supply chain. These companies—the OEMs and their EMS partners—need to comply with certain product standards. There are limits on what they can use, for example, compliance and standards regarding the use of lead (Pb).

However, to get the level of detail needed—the percentage of that substance actually in an approved component and how that adds up across a device—is a lot of data. Consider the granularity of combing through databases, retrieving data, verifying it, and cross-checking it against all regulations.

New substances are added to the database every six months as science and testing advance, making it even more complicated. So, after completing the above processes and getting your data, the same manual effort must be repeated each time a substance is added to the database.

There are high costs associated with this data mining and verification work. As your regulations increase, your compliance costs also increase. Many of these companies have compliance teams that are overwhelmed.

But we’re not just talking about regulations and compliance. It now also touches on supply chain resilience, because people also want to know about a substance or component from a critical-rare-earth-minerals perspective.

Finally, you have the trade compliance angle, but it all boils down to the same thing: You need to know what's in your components, where they are coming from, and how they're made. This is a primary and significant problem for OEMs and EMS companies, and it's causing higher costs and lower productivity.

LaRont: Compliance costs have always been a major operating factor in manufacturing. So, anything that speaks directly to reducing costs and complexity is a significant value add. How does Banyan.eco use AI and LLMs?

D'Souza: Large language models are based on prediction models at a very simplistic level, meaning they predict based on what's come before. This is very good for understanding context and generating stuff. Generative AI is all about that. You want to generate a sentence? The AI knows what should be next in your sentence based on what you begin writing. This type of AI is also very good at looking at and gathering data.

But it does not work well when applied to very specific use cases, such as deciding whether something is compliant, especially when you have many rules that are highly sector specific. AI is not at the level it needs to be to make specific decisions on industry and domain-specific topics.

That's what we are trying to do. We have figured out how to use AI for the first part—the time-consuming, cumbersome data gathering.

Then, how do we ensure that the decisions made, whether by AI alone or by AI augmented by external inputs (e.g., human intervention), are correct? Manufacturers need tools to perform these tasks at the level of a compliance expert, so we are working to refine the AI for increased productivity.

LaRont: Tell me about the process and how your tool works.

D'Souza: First, we don't just use LLMs. Let’s use the example of a BOM. The LLM is instructed to obtain data for each component from the manufacturer’s datasheets for data acquisition. The LLM retrieves this data and then verifies it for a particular data point. An LLM can do that. This is like a compliance manager in a company visiting a manufacturer's website to find the data. It's very restricted, and occurs in a very legalistic framework.

The next step is more interesting. Let’s say you have a BOM with 300 components, so the data is being pulled for all of them. Now, this data has to be crunched, digested, and totaled to create a compliance report for the device as a whole. This is a secondary level where decisions are made. Now we are getting into what we started discussing at PEDC and will continue discussing in our white paper at APEX EXPO.

We have a deterministic layer to ensure that there can be no hallucination, that nothing is made up, and, further, that there is mathematical proof of every outcome. I just got off a call with someone who, last year, told me that he had planned to use ChatGPT and a prompt engineer to get his company’s compliance information. On the call, he said they stopped doing that because the information they were getting had to be reverified on the manufacturer’s website, and there was so much inaccuracy; false facts were being given. An LLM that gathers your data with a deterministic layer providing mathematical proof can give you that productivity boost. His company has just become a customer and will start using the platform in April.

LaRont: That's a great story. The issue of hallucinations is real. As we all use AI in more parts of our lives, we give it confidence that we really shouldn't.

D'Souza: Exactly. There's a thing called mathematical proving or generative theorem proving (GTP). It provides a kind of certificate at almost every step of the process that proves, “I’ve reached this decision because of… “ Now, if someone wants to audit, the whole decision tree is clearly documented and available for review.

LaRont: You have mentioned productivity gains. What are the productivity gains from using a tool like this versus the way that we've done it?

D'Souza: The biggest one is that your employees can get a lot more done. You no longer need to massively overload your team with data mining and processing. This is a time productivity gain of at least 50%.

Also, the databases used today are static and outdated. Someone must enter the data to update these databases. There could be thousands of people working on that. But why do we need to rely on that static database when, with AI agents, you can now have live in real-time data?

LaRont: That is tremendously valuable.

D'Souza: Because of the power of AI, you can also build and automate the whole workflow, end-to-end. One set of AI agents gathers the data, another applies mathematical proofs, and a third generates a final report in the template selected by the company.

All the compliance manager does is upload the bill of materials, and later, depending on the size of the BOM, the report will be ready to send to the customer. The time that is spent for a single compliance instance is cut by at least 50%, and that is understating it.

LaRont: That sounds like a positive contributor to job satisfaction as well.

D'Souza: Yes, more importantly than just saving time, these are knowledge workers. They are talented, and their intellectual capacity should be used for more important things, like how to make a better, more sustainable design, or how to design for circularity and not have to use a part that is toxic or depends on rare earth mineral availability. That uplevels a company’s productivity.

LaRont: We certainly need our finest engineering minds working on these larger issues. Francis, tell me about your upcoming presentation at APEX EXPO.

D'Souza: As I’ve mentioned, we built this deterministic, mathematical layer, and we benchmarked it with other AI engines. We tested the big household names today, including Anthropic and OpenAI, and retested them to see how they handle a BOM for compliance. What data are they getting back? Are they hallucinating, etc.?

We put our construction against theirs using our AI-deterministic model. Through benchmarking, we set a threshold to define the number of false compliances produced by AI tools, and we found that, for the most famous AIs on the market, the false compliance rate ranges from 8% to 13%, which is very risky. Companies have full liability for any inaccurate or false declarations they make. In contrast, with our approach using the deterministic layer, we found no instances of hallucination or false compliance, zero, essentially no risk.

We will be presenting our findings at noon Wednesday at the Technology Pavilion on the show floor. We will be explaining how to actually harness the power of AI, because we do use AI in our solution. We were able to show how our solution was better than any of the big AIs. They are very powerful but are general-purpose large language models. What we have done is build a model specific to this use case for the electronics industry, and that's a difference.

LaRont: Francis, this has been really interesting. Congratulations on your Rising Star Award, and I wish you luck with your white paper presentation.

D'Souza: Thank you, Marcy.