Founders in Mobile Co-Founder, Ivan Crewkov, on Building AI for Language Learning

Ivan Krewkov

Is it possible to achieve fluency in English without ever living in an English-speaking region? Or will learning English continue to be reserved for those who can afford the best tutors? co-founder, Ivan Crewkov, knew the answers to these questions could open up opportunities for millions of language learners globally. A virtual AI-powered robot that teaches kids English, is the first of its kind in EdTech, an area that investors are increasingly turning their attention toward.

In this month’s Founder Profile, Ivan shares his approach to building an entirely new product through rigorous testing, a focus on monetization, and a personal quest to make language learning accessible.

On a mission to democratize English speaking practice

My children sparked the idea. I had recently moved to the Bay Area from Siberia with my family, when my daughter was only a few years old. We speak Russian at home, but I wanted my daughter to also be fluent in English. I discovered that outside of her pre-school, where she could be immersed in English, there actually aren’t many opportunities to practice speaking English. Although there are many ESL websites and products, they tend to focus primarily on reading and vocabulary.

The only option to practice speaking in a digital environment is with a human tutor over video conferencing. This is very expensive, costing about $300 per lesson. The vast majority of families in the U.S. can’t afford this. The children who need this most tend to come from immigrant families that can’t afford this. And, around the world, half a billion children try to learn English as a second language every day. Only a tiny fraction will ever reach true fluency—partly because of the lack of speaking practice. Classes are large and teachers are overwhelmed everywhere, so even though English is a mandatory school subject worldwide, many, many children don’t ever learn to speak.

I had started recording lessons that my daughter was taking with human teachers, and realized that 80% of what human tutors do over video conferencing could be done with voice technology and voice assistance. Prior to working on, I had been a co-founder and CMO of, where we were building an Amazon smart speaker before Amazon built Echo. Because of this background, I knew we could reapply voice technology and voice assistance to education and ESL.

The challenge of automating language learning

The hardest part about building in this space is that it’s hard to automate ESL while explaining complex concepts. Much of ESL teaching is dedicated to routine things like vocabulary and drilling simple language mechanisms, and it’s hard to keep children engaged.

To solve this, we built a virtual, voice-first character that uses the same approach and same mechanics as a human tutor, but with additional game design. This works well for children, because for them, it’s like playing video games.

Another big challenge is being the first in this field. We are the first virtual tutors of its kind, and there are not a lot of examples to learn from. AI is not sci-fi. It’s a code that you need to write and collect, and we had to create everything for the first time. We started by learning from real tutors and teachers, and identifying what we could directly apply from them.

Be efficient with your funding in EdTech

Initially, it was difficult for us to raise funds. Before the COVID-19 pandemic, EdTech wasn’t attractive to investors—they didn’t want to mess with products with children. It can be a challenging industry, because you have to sell a product, and the users are children, but the buyers are parents. So you have to know how to sell twice. Investor interest changed after COVID-19 when EdTech became necessary for learning.

Given our initial challenge with fundraising, we had to be efficient from Day 1. We learned to generate revenue and cross-profits. In those early days, Braavo was really helpful. When you’re starting out, you need that cash in your hands to run your business. But even when you start to generate revenue, it takes too long to get your earnings back from the app stores. Because of Braavo, we were able to pay for essential business functions, and then began to scale and generate more and more revenue. In 2019, we grew 10x year-over-year. As we started to grow faster, investors began to notice us, and we were able to raise more funding.

Defining what a successful learning session means

I believe in growing with what you’ve got and monetizing as soon as possible. While we want to raise as much as we can, I know from prior experience that it’s very stressful to run a business without generating revenue or even user feedback. So for, we ran as lean as possible from the start, and we still try to be as efficient as possible.

LTV is our primary business metric, as we are still in the product development stage. On the engagement side, we measure weekly and monthly retention. Our overall North Star metric is our unique customer learning success metric: When the user completes a successful voice interaction with our virtual character. That’s because each time the user interacts with Buddy, the user has a chance to learn something. So each completed interaction helps indicate completion of a unit of learning.


We test everything. We test visual styles, paywalls, learning exercises, changes to learning mechanics, everything. We don’t release anything without A/B testing.

We also collect a lot of data from every student’s interaction with the app and during their learning sessions. On average, we’ve found that our students spend time learning about three times per week, and each session lasts about 25 minutes. We track how students use voice assistance and how much, and which sessions become longer sessions.

A subscription-based revenue model that considers parents

We charge users early, and our weekly model works as a paid trial. We find that users unsubscribe from the weekly plan and switch to annual to save money. We don’t have free trials, although we do have a very limited free content plan.

To get here, we tested a lot of crazy ideas. Eventually, we found a combination that works for us. For instance, in the past year, we charged $7/week, or alternatively $35/year. We found that the majority of users pay for the annual subscription. They feel comfortable in the app, and want to take their time (and save money) to figure out the best way to use it, without the weekly payment looming over them. These upfront payments, with users pre-paying for long-term plans, covers our CAC.

Our particular subscription model works because we have two types of users: Children, and parents, who don’t have a lot of time to think about small purchase decisions. Some parents will decide to pay for an educational app before installing, perhaps because the app was recommended by their friends. But if a parent finds the app through paid user acquisition, you need to convert the parent as soon as possible. Otherwise, you’ll lose them: They’ll postpone the purchase decision, thinking they’ll give the child time, like over the weekend, to try the app for free.

I had once heard that parents spend on average seven minutes on the app with their children when they install it. So that purchase decision needs to happen within 7 minutes. That’s how we thought about creating a subscription model for our app. We did experiment with different things and a lot of free content before we arrived at the monetization model that works for us.

Solve problems that are personal to you, and start monetizing

Understand the problem you’re trying to solve. The problem should be personal to you—ideally, you can personally test your product every day.

On the business side, start charging your users ASAP. It’s important to know if your product is affordable or not. Paying users are the source of the most valuable feedback for your product.

Security check
Please complete the reCAPTCHA below to proceed.