We started working on Humbot full-time in June 2017. Even though we started with a different product, we always had the same goal: Improving the public understanding of science.
Here is the summary:
In November 2017, we realized that there’s an opportunity in the tutoring market. Now, we’re offering online tutoring 100 times cheaper than existing competitors. We tutor our students through a familiar medium, Facebook Messenger, that has low barriers to entry thanks to its asynchronous nature. Just like ridesharing companies, we can provide the service at these prices because we are confident that we will be able to drive down our costs significantly within the next two years through automation of the tutoring.
Here is the longer, more fun version:
At the beginning of a startup, you need to find an angle to get started. A way to take steps towards your end goal that also allows you to monetize and gain users quickly. Finding this angle is not apparent, but the result of a month-long trial and error process. In general, the more specific you can be about your product and target market, the better. The reason is that way you can quickly test your hypothesis, see it fail, and move on to the next iteration of it.
Struggling to gain traction and YC
When we started, Humbot was a native app where you could read short, simple explanations of scientific topics. In August 2017, we had decided that users should be able to ask questions and that we, as experts, can answer them. A month later, in September, we started to focus on the question answering exclusively. As a consequence, we scraped our iPhone, and Android apps, relaunched our website, and started a Messenger bot as our new and only product offering.
We started looking into ways of how to do question answering with machine learning. It turns out that, in recent months, there has been good progress in research in this area. Even though it’s not possible to do automatic question answering yet, we concluded that it would be within the next two years. At that point, we were doing science question answering on Messenger with a semi-automated bot, hoping that we can automate most of it as soon as possible. It seemed like a cool product: A friend in your pocket that answers all your science questions immediately.
After a few weeks, we still didn’t have any real traction. We still decided to apply to Y Combinator. Unsurprisingly, we received our rejection on the 24th of October, 2016. “Thanks for applying to Y Combinator, blah blah blah”. Eager for more feedback, I wrote a message to Tim Brady, one of the partners of YC.
Surprisingly, he replied swiftly:
After a few generic sentences, the last sentence is what made a significant impact on our way of thinking: “Google search sets a pretty high bar as an alternative”. What the shit?! Why does he think we’re competing with Google?
Our product is too broad, that’s why! Science question answering? Come on, that’s precisely what Google does (among other things).
A clear focus helps to iterate faster
We decided to focus this bad boy. The first thing we narrowed was our target group and use case. This would help us make our marketing, communication, and product offering crisper. It was evident to us to pick high school students as our target group. They must have questions on school-related stuff like math and science daily.
The next step was to understand how to get in front of the kids. One approach we tried is to cold email teachers. A tool they could give to their students. Didn’t work out. Out of the 100 or so teachers we reached out to, a single one wanted to give a try but then decided not to, because he didn’t want to use his personal Facebook account.
The next thing we tried worked. I don’t remember exactly how we came up with the idea, but we thought that we should have a look at the tutoring market. It turns out that the tutoring market is large, growing, dominated by old-school incumbents and high prices. A perfect situation for a startup to shake things up.
Since then, our weekly active users grew 50% each week.
Not only that, but our retention and engagement rates have also skyrocketed. Some users spend a couple of hours per day on Humbot (of course not every day).
An old-school, manual market with high prices
I don’t want to deep dive into the tutoring market here, but what you need to understand is that parents are paying exorbitant amounts of money to people tutoring their children. A lot of the tutoring happens online through the companies’ custom built, Skype-like video calling software. Hourly rates go up to a couple of hundred dollars, with the average being around $70 per hour.
This showed us two things. First of all, kids struggling at high school seems to be a massive problem (because of the enormous amounts paid). Second of all, these high prices are a consequence of humans sitting in front of a webcam and explaining the same things over and over again. A market that is marked by high rates due to apparent inefficiencies and a problem that is gigantic? Count us in!
We’re now offering tutoring for high school students over Messenger. Although some of the interactions are automated, the actual tutoring is done by real humans. Students pay $0.20 per explanation they get. Some explanations take half an hour of a tutor’s time. That’s literally 100 times cheaper than existing online tutoring solutions.
Not only that, but it’s also a much more natural way for teenagers to interact with other humans. It’s incredibly low barrier: You just start texting when you have time. Often, they snap a picture of their problems, draw something on it and we write or draw something back. Some explanations may span days because a student might start with a question, but then wouldn’t have time to respond for a day or two.
In addition to the price, the familiar and asynchronous nature of our service is also a better fit with this target group.
How can we provide tutoring in such an affordable manner? Because we believe that we can automate most of the tutoring in the next two years. Therefore, it is worth investing time and money now.
It will be a gradual process. There probably won’t be one single machine learning model that does everything. Instead, we will start automating step by step. Requiring less and less human supervision.
Not only will this be more efficient for us, but it will also improve the product’s value. Students will be able to get help instantly and 24/7. Furthermore, we can provide the highest quality of tutoring to everyone, independent of the skill levels of every single tutor. We expect another added benefit to be the high value we can extract from our data: For example, to see what kind of explanations work better or worse.
A good analogy for our strategy is ridesharing companies like Lyft. They are not at all profitable right now, but they know it’s worth investing because once they have mostly self-driving cars, the economics will work out.
I know that we’re just at the beginning and a lot of things can go wrong. But what drives us (pun not intended) is to create a new kind of school. A school where teachers can focus on the fun, non-repetitive part of the profession and instead of teaching 30 students, can teach 300. We think that we can build the software that gives teachers, and with that to our future generations, superpowers.
If you’re an educator and like what we do, contact me (firstname.lastname@example.org), so we can explore ways to work together.