Our guest for today is Andrew Will Skrypnyk, CEO and co-founder of Promova, a leading platform for language learning. A Forbes 30 Under 30 honoree, Andrew is a serial entrepreneur with a background in software engineering and a passion for building sustainable, forward-looking businesses

In this episode, Andrew shares what early AI workflows looked like before LLMs, why agents represent the next leap forward, and how Promova replaced human support with AI while improving customer experience. He also discusses the challenges of driving AI adoption among non-technical teams, and why leaders must model usage themselves.

---

About Andrew: LinkedIn | Instagram | Facebook
About Promova: Download the app | Instagram

ABOUT ROCKETSHIP HQ: Website | LinkedIn | Newsletter | Youtube | Podcast Website

FULL TRANSCRIPT BELOW
Shamanth Rao:
I’m excited to welcome Andrew Skrypnyk to Intelligent Artifice. Andrew, welcome to the show.

Andrew Skrypnyk:
Hi, Shamanth. I’m so excited to be here—not just because of the impressive work your team does, which people can see on the Meta Ads Library, and which I’ve certainly followed, but also because of your journey to this point. From the very early days of AI to building an AI-first company, I think what you have to share offers great reflections on the future of work, the future of AI, and more. I certainly learned a lot even during our prep call, and I’m excited to hear what you’ll share today.

Shamanth Rao:
Wonderful. Let’s dive in. For a lot of people, “AI” today means LLMs. But your team was working with automation and AI long before LLMs appeared. What did some of those early workflows look like? Understandably you didn’t call it “AI” at the time—but what carried over into today?

Andrew Skrypnyk:
You’re completely right. Today we call AI everything from machine learning systems to automations and workflows. Even some SaaS automation companies—basically just running scripts and algorithms—now call themselves “AI automation.” In that sense, AI has been with us for decades.

At Promova, we used AI from the very beginning to automate processes, campaigns, and product features. For example, we built a pronunciation practice feature, where learners could improve their syllables and sounds. That was machine learning-driven. We trained our own model and integrated it into the app. Back then, we just called it an “ML module.” Today, it’s labeled “AI.”

We also used AI in the back office—for operations, video generation, and image generation. I can’t think of a single part of the company that doesn’t use AI in some form.

Of course, with LLMs, we’ve taken things to another level. For example, we now use large models to create new experiences, such as role-play lessons where learners can practice English in writing or speaking with a bot, almost like real dialogues. We’ve also automated support operations, content generation, and personalized content recommendations. By feeding smaller models with user interests (not private data), we can generate lessons and curricula tailored to them.

So in the last couple of years, AI has pushed us to a new level.

Shamanth Rao:
That’s fascinating. You started this long before LLMs became mainstream. You mentioned building ML models for pronunciation practice—can you share more about how those worked? Today, LLMs make it look much easier.

Andrew Skrypnyk:
Yes. LLMs are chat- or voice-based tools. They predict text sequences, but they don’t analyze the quality of sound. Our ML model did exactly that—it assessed pronunciation by analyzing sound waves.

We built models not just for English but also Spanish, German, and French. For English, we trained on multiple accents—British, American, Australian, and others. Even today, companies like ours still use ML models for specialized needs such as pronunciation assessment, because LLMs can’t handle that specific use case.

Shamanth Rao:
That makes sense. I’ve seen Python libraries that evaluate sound quality—it sounds like what you built was a much more advanced version. I’d also assume that using LLMs for this would be far more expensive in terms of API credits.

Andrew Skrypnyk:
Honestly, I never tested LLMs for pronunciation assessment. My team may have experimented, but I don’t believe LLMs can do that. It’s not a cost issue—it’s a capability issue. They simply aren’t designed for it.

Shamanth Rao:
Got it. Let’s switch gears. You were using ML-based models years before ChatGPT. What was your first experience with GPT like? What struck you about its potential?

Andrew Skrypnyk:
The first time I heard about GPT-3 was around 2017. A friend had access to an early beta and showed me how it worked. Even then, it was mind-blowing.

When GPT-3.5 was released publicly around late 2021, one of the first serious prompts I gave it was: “Act as the CEO of my company. Here’s our stage, product, and competitive landscape. Build a growth strategy for next year.”

It came back with a surprisingly solid plan. Some ideas were obvious—low-hanging fruit—but others offered fresh perspectives. I actually used that strategy that year. It was like having someone in the room to brainstorm with.

That was a breakthrough moment for me. I realized this could serve as a mirror—helping me refine ideas and see things from another angle.

Shamanth Rao:
Yes, that resonates. I’ve spoken with people who use GPT almost like a therapist. It doesn’t necessarily give “mind-blowing” ideas, but it reflects your thinking back to you, which is clarifying.

Andrew Skrypnyk:
Exactly. We often expect LLMs to be “superhuman,” but they’re limited by the data they’re trained on—essentially the internet. What they can do very well is simulate other perspectives.

You can prompt it to act as a marketing director, a user persona, or even a competitor CEO. That gives you new opinions and viewpoints you may never have considered.

Where things get even more powerful is with agents. I recently used Comet as a personal assistant—it helped me find an apartment, research locations, and even plan around San Francisco by combining tools like Google Sheets, search, and maps. That’s where real value lies: using AI as a creative tool for automation.

Shamanth Rao:
That’s incredible. And with multimodal tools, the possibilities expand even more.

Andrew Skrypnyk:
Yes. For example, I used AI to find inspiring places to work around San Francisco with a nice view and warm climate. It automatically mapped out locations for me. Apps are being built on top of these models, tailored to user needs—just like the App Store boom from 2008–2018. I think we’re heading into a similar golden era for AI-powered apps.

Shamanth Rao:
Let’s talk about your product features. You mentioned some, like pronunciation, still rely on ML rather than LLMs. How do you decide what should use AI, what should use ML, and what should just use code?

Andrew Skrypnyk:
Great question. My background is in software engineering, so I naturally think about efficiency. For basic quizzes—like translation or multiple-choice—we just use code. No need for ML.

For more complex tasks—like pronunciation—we use ML models. And for text-based tasks—like grammar correction, writing practice, or dialogue simulations—we use LLMs.

We ideate internally, look at competitors, run fake-door tests, release MVPs, gather feedback, and iterate. That’s the cycle. Users often surprise us by using features in unexpected ways, which guides our next steps.

Shamanth Rao:
You also mentioned automating customer support. What was that experience like?

Andrew Skrypnyk:
That was fascinating. Initially, we had only human support—24/7. Agents often used scripts, so some users felt like they were talking to bots.

When we switched to AI-based support, powered by our knowledge base and rules, user satisfaction actually improved. Fewer people complained about “not talking to a human,” and more people thanked us for helpful, empathetic responses.

For the business, it reduced costs, resolved more tickets, and even improved revenue retention. AI-generated answers were simply more personalized and empathetic than scripted human ones.

Shamanth Rao:
That’s amazing. And it shows how AI can empower teams rather than just replace them. How do you ensure non-technical team members also adopt AI in their day-to-day work?

Andrew Skrypnyk:
That’s been a challenge. We weren’t pushing hard enough last year, and now we’re catching up.

The key is leading by example. Showcasing real use cases. Appointing someone in the company to champion AI adoption. And most importantly, educating employees—just as we once learned Excel in school, employees today must learn how to use AI tools.

If we don’t, we’ll be left behind.

Shamanth Rao:
Yes, leadership has to set the tone. Andrew, this has been incredibly wide-ranging. I’ve learned so much about how you transformed your company from pre-LLM to post-LLM, and how you’re applying AI across every function.

Before we wrap up, can you share how people can find you and your work?

Andrew Skrypnyk:
Thank you, Shamanth. People can find me under “Andrew Skrypnyk” on Instagram, Facebook, LinkedIn, or by email at [email protected].

Our company, Promova, is a language learning platform helping people worldwide. Outside of work, I’m active in extreme sports—kite surfing, snowboarding, Ironman training—and I also love golf. If anyone in San Francisco wants to join me, let’s do it.

Shamanth Rao:
Wonderful. Thank you for your time, Andrew. I’m excited to release this to the world.

Andrew Skrypnyk:
Thank you, Shamanth.