Panos Stravopodis CTO and Co-founder at Elyos AI, brings two decades of software engineering and a passion for operational transformation to the clean energy and field services world. Elyos AI builds vertical-specific AI agents - helping companies accelerate project delivery, optimize resources, and streamline booking and communications, especially for fast-moving businesses in energy and field services.

Listen to the episode on Spotify, Apple Podcast, Podcast addicts, Castbox. You can also watch this episode on YouTube.

In this episode of the AI Minds Podcast, Demetrios talks with Panos about Elyos AI's path - from a Y Combinator-powered sprint to what it takes to spot the real operational bottlenecks in clean energy.

Panos discusses his progression from building scrappy reservation software to running engineering teams in Amsterdam and London, and the pivotal move that started Elyos AI’s journey. Hear how the initial product, focused on energy optimization for commercial properties, shifted as the team discovered that slow, fragmented communication—not software or hardware—was the true barrier to large-scale decarbonization.

The conversation dives into how Elyos AI’s AI agents take on the high-volume, repetitive workfield—mainly booking appointments and managing calls so small, hands-on teams can focus on bigger business priorities. Panos highlights the “network effect” among trades businesses, the unique trust-building challenges in these markets, and why staying vertical leads to deeper impact. He also explains the onboarding process, balancing deep customer understanding with scalable automation, and how agent performance is measured relative to real baseline operational data.

Listeners will learn about:

• The unglamorous but critical pain points in field service and clean energy implementation.

• How rapid prototyping and customer feedback led to a significant product pivot.

• The role of trust and specificity when introducing AI to traditionally underserved verticals.

• Pano’s view on the future: how focusing on core expertise trumps premature over-generalization in AI for operational automation.

Show Notes:

00:00 Startup Journey: Tech Roots & Early Projects
06:40 Y Combinator & Team Relocation
14:15 Early Focus: Energy Flexibility Optimization
21:05 Discovering the Communication Bottleneck
28:03 Building AI Voice Agents for Real-World Operations
34:55 80/20 of Field Service Calls: Bookings & Scheduling
41:13 Network Effects and Community Building
45:01 Why Vertical Focus Wins—For Now
52:00 Onboarding, Baseline Data, & Success Metrics

More Quotes from Panos:

Demetrios:
Welcome back to the AI Minds podcast. This is a podcast where we explore the companies of tomorrow being built with AI first. I'm your host, Demetrios, and this episode, like every episode, is brought to you by Deepgram—the number one speech-to-text, text-to-speech, and now, voice API on the internet today. Trusted by the world's top conversational AI leaders, startups, and enterprises, some of which you may have heard of, like Spotify, Twilio, NASA, and Citibank. I have the pleasure of being joined today by the CTO and Co-founder of Elyos AI, Panos Stravopodis. How are you doing today, man?

Panos:
I'm very good, man, and thank you for having me here. Pleasure.

Demetrios:
So, I'm excited to chat with you about a bit of your journey and how you went through Y Combinator and eventually pivoted, and what that took for you to move into the space that you're in right now. I know you saw pain in the market when you eventually landed on what you did—I'm almost teasing what you landed on right now. But let's start with how you got into tech.

Panos:
Oh, that takes us quite far back. I got into tech the first time working for a small boutique hotel in Zakynthos. They were looking to build a new reservation system. That was back in 2005, very early days, there were no platforms ready to go. I always liked computers and was into coding, but was very Microsoft-oriented at the time, like Visual Basic and such. I started building the platform, and I found it quite interesting. I thought, "That's definitely what I want to do for the next few years."
I went to university after that, did my bachelor’s and master’s in computer science. Then I moved to Amsterdam to do my software engineering degree and started working in Amsterdam, and eventually in London. So, the biggest part of my background is software engineering and building cool things in different sectors.

Demetrios:
So what gave you the audacity to think that you could build your own company?

Panos:
I think it was more about having the control to decide what the big problem you're solving is. Also, how empowering that is—you work with your team to solve a big problem that you define. You have the opportunity to tackle some of the biggest problems out there, and you can move as fast as you can. That’s very powerful. When I realized we’re putting in the hours anyway, I thought: why not go all the way and build the best thing for tomorrow? Let’s make it happen.

Demetrios:
When you first set out on your entrepreneurial journey, what did you have in mind?

Panos:
I was very excited because the sky's the limit in terms of ideas. We knew how to build things since I already had about 10-15 years of experience. But I hadn't previously founded a company; I'd fundraised before as part of previous roles, but it's not the same, especially when you start from zero. It’s all about the founding team, identifying the problem you’re solving, how you'll solve it, and then gaining trust from both investors and the first hires—people who are joining you at that early stage when everything is new. There was definitely excitement but also a lot of unknowns. I’m really good at uncertain situations. I love figuring things out and being in a scenario where you’re building something when you don’t know what’s coming in two weeks, a month, or five months. That’s exactly how the AI space is right now, and it’s why I’m very excited to be building in that space.

Demetrios:
So what was your first idea?

Panos:
We did Y Combinator in the summer of 2023. Pip, Adrian, and I got together at the beginning of 2023, started talking about a bunch of ideas. Pip and I were in the energy space at that point and were excited about opportunities, especially around optimizing energy consumption and demand response programs, as well as energy flexibility in general. Adrian also had a background in energy from his time at BCG. We thought, "Let’s apply to YC—put a couple of great ideas in the application and see how it goes. We definitely wanted to do YC, so we decided to go for it."

We heard back at the beginning of May; the batch started in June. In May, we got an email saying, "We want to talk to you, let's jump on a call." Long story short, 24 hours later we were accepted. We had about a month to sort out our lives. Pip and I were in London and Adrian was in Singapore (his previous company was there). We all had to move to San Francisco within a month, which was very intense, but we made it happen. I think it was the best decision of our lives at that point.

We went to San Francisco, spent the summer there, and started building in the energy space, framing it as AI agents for commercial and industrial buildings—trying to optimize their energy consumption and figure out opportunities for them to participate in energy flexibility programs.
If people aren't familiar, energy flexibility is when all the different assets in your building can take part in energy programs, where you can sell energy back to the grid either by pausing use or, if you have batteries or other assets, exporting energy back to the grid.

We did YC and came back around mid-September 2023. We were in the process of building the product and started hiring engineers—first Alex, then Matt. For the next year, we built in that direction. But as we worked more closely with customers, we realized most were anxious to work on their decarbonization projects, some with very aggressive targets. But for various reasons, these projects were delayed.

So, we dug into what was causing delays. For us to enable asset participation in managed flexibility, solar panels needed to be installed. But the real struggle was that things were really slow at the start of the project—before even installations or planning permissions and such. We worked with customers to understand what was going on and realized a lot of people involved in the process, especially for big solar projects, needed planning permissions, insurance considerations, procurement teams, and installers to all coordinate. Most of these people communicated by phone calls, with some emails, but there were long lags between messages—sometimes two to three days, sometimes two weeks, depending on their schedules. They’re always on the go, working on different projects, attending multiple locations—it becomes complicated.

Demetrios:
So when you say "all these people," are you talking about people implementing the project for the company, the people hired to install the panels, or both?

Panos:
All of them. But in particular, the people we worked most closely with—the installers or maintainers of the buildings.

Demetrios:
I see.

Panos:
We quickly realized this was the main blocker for customers to deliver these projects. So, we connected with partners to understand the deep-rooted problems and what they were trying to do. About a year ago, voice technology was just starting to become a thing. Before summer 2024, models were not so conversational, but Deepgram and other solutions were beginning to mature and pipelines for streaming were becoming viable.

As we started working with customers, we found the biggest burden was how they run operations internally—especially around their call centers. They needed agents to receive incoming calls and make lots of outbound calls to confirm appointments, order parts, or coordinate with planners for larger projects—lots of scheduling required.
We built a prototype in a few weeks and started testing it. Even as a simple MVP, people immediately saw the value. We started with things customers would otherwise miss—like after-hours calls or overflow they couldn't handle.

The response was really positive, and we saw a network effect: in trades businesses, everyone knows everyone, so happy customers recommended us to others. We built core integrations so that our agents could connect with the major CRMs—so the agent can book appointments, take payments, handle scheduling, and other major workflows via phone.

Demetrios:
Is there an "80/20" to what these phone calls are about? Is it mainly scheduling or confirmation?

Panos:
The vast majority is booking appointments. LLMs are great for providing context for general inquiries—"Do you do this? How do you help?"—but by the time a customer calls, they're ready to book. Maybe they don’t know exactly what brand they want to install, but they know whether it’s a boiler, solar panels, etc. Our agents (and the human installers on first site visits) can help with those decisions.

I would say 80% is bookings, and also managing bookings—customers calling back to reschedule. These are the majority of calls. However, businesses have bigger internal problems that get postponed because they're busy handling these calls. Our agents help them focus on more important things and grow their business—getting time back, increasing bookings, and handling more inquiries.

Demetrios:
I like how you mentioned the network effect. This market—installers, field services—is huge but underserved, right? But that probably means those who do serve it are loved.

Panos:
Absolutely right. It's definitely an underserved market, but there’s a reason for that. Many of these businesses are very hands-on—a small office team, founders and managing directors deeply involved in day-to-day operations. Introducing new technology requires building a lot of trust, especially since the tools we're providing weren't even technically possible last year. These businesses also tend to be very complex from an operations perspective—who are their customers, how do they serve them, what range of areas do they cover, etc. That complexity may help explain why this segment is underserved.

We work really closely with our customers, and we see firsthand how their daily operations change. What we're building is more like an operating system for the AI era, not just an AI agent you sign up for and use generically. We deeply understand each customer’s operations and pain points and focus on delivering maximum value for their way of working.

Demetrios:
Do you see this as something you could generalize to other fields, or would you rather go super deep on the energy sector?

Panos:
It's a good question. For now, we want to stay focused on our vertical. The models today do very well when a problem is clearly defined. We have a strong understanding of our space—trades and field services. To serve other industries well, you need that same deep customer understanding. As technology improves, generalization will get easier, but for now, going deep creates the most value. If you look at CRMs, technically they could be generalized, but you usually see specialized CRMs succeed in specific sectors—trades, healthcare, marketing, etc.—because that specialization is what customers need. Our priority is to stay focused.

Demetrios:
Talk me through onboarding. What do new customers give you so you can add value?

Panos:
We start with recordings of previous calls. To build effective agents, it’s important to understand how calls are handled. We look for baselines: percentage of queries resolved on the first call, how many calls a customer needs to make to resolve an issue, etc. Bigger companies might already have some of that data, but smaller teams often don’t. As part of our onboarding, we help create those benchmarks, which also makes it easier to score the performance of our agents in comparison to human agents.

We spend time understanding the business, which is the most important thing for us—not just deploying an agent, but understanding the workflows, helping streamline processes, or recommending improvements. We also collect company knowledge: process documentation, FAQs, internal procedures for things like boiler installations, and more. We use these to prompt and build custom evaluation sets. We even select the best agent voices based on sentiment analysis of previous calls, so our agents reflect the best performing human call agents and simulate their behavior as closely as possible.