rey.ai: The meandering journey of building a startup
A series of successes, failures and hard lessons
When I read historical accounts of startups, the stories often fall in one of two categories:
- Flawlessly executed companies or,
- Epic failures, where company xyz crashed and burned.
The problem is that both narratives present a false dichotomy of a far more nuanced reality. Many startups lie on a spectrum somewhere between these binary outcomes.
This is the story of rey.ai, a startup whose fate fell in the middle of the drastic unrelatable outcomes mentioned above. We’ve learned a lot in the past eight months, and we would like to share some of these learnings with the rest of the community.
A secondary motivation for this post is to address the underlying reasons that led us to put the business on hold. We are extremely grateful to our friends, family, advisors and everyone who contributed to our journey, and therefore owe them an explanation.
Some context
Our first idea was to build an A.I. to screen resumes and surface the highest quality candidates, saving hiring managers hours of exhaustive work. The tool would enable them to focus on qualified candidates, instead of countless hours sifting through resumes.
It was a pain point felt by many in the field, particularly at large companies where job postings receive 100s if not 1000s of applicants. My partner Tom had already been working on the machine learning model as a side project, so we even had a rough prototype built.
But when we started looking at the market and talking to people, it became clear we were a few years too late. Several companies had already launched products, and admittedly they were doing it well. If we pursued this path, we would be undifferentiated, late to market, and significantly undercapitalized.
We hadn’t even incorporated and it was already time to dispose of our first idea.
While brainstorming a new direction, we thought back to how we met: a serendipitous introduction through a mutual friend. A year later, we would quit our jobs and co-found a company. That single introduction changed the trajectory of our lives.
For us, this sparked a fundamental question: at a global scale, how many valuable relationships should exist but don’t because there isn’t a mechanism to make them?
We’ve already seen A.I. begin to automate certain human-driven tasks. Naturally, it felt reasonable to assume that A.I. could be applied to the task of networking. We believed with the right inputs, machine learning could computationally predict valuable relationships, and then use email to facilitate them.
This idea stuck, and no one was directly competing with a similar product.
We jokingly referred to Martin, the friend who introduced us, as Marty.ai, because that single, double opt-in email introduction would serve as an inspiration for Rey and its product experience.
The problem we were solving
The people you meet have a profound impact on your career. To borrow from Porter Gale:
“your network is your net worth”
Networking is particularly important in industries such as recruitment, startup fundraising, mentorship and sales.
Today, professional networking is driven by a combination of human effort and serendipity. As a result, meeting the right person is often time consuming, expensive or random. Recruiters cost tens of thousands of dollars, cold messages are unproductive, and events take days out of your schedule.
That was our framing of the problem, and we spent the next 8 months building Rey to address it.
Our minimum viable product
Rey’s job was to automate the process of professional networking. It had three main parts:
- A web app to gather info from users
- A deep learning engine to predict which introductions are worthwhile
- An email service to make the introduction after both users opt in
Rey would work almost entirely in the background, obsessing over who should meet. Human effort was reduced to just an email reply, and Rey would get smarter with more users, producing powerful network effects over time.
Assumptions we made along the way
We used lean principles to guide us towards product-market fit. We constantly tested our assumptions and would rapidly course correct as we learned. We didn’t quite get to product-market fit, but we had a fun ride and learned some hard lessons along the way. Here they are:
Assumption 1: Resumes are the right tool for networking
We were totally off here, but we recognized it early enough. Resumes presented a lot of sign-up friction. People were confused by the relationship between networking and resumes. We’d often hear, “resumes are for job hunting”. Users often wouldn’t even open other users’ resumes before opting into introductions, and the A.I. had a hard time parsing the relevant info.
We got rid of resumes in favour of “tweet sized bios”, and subsequently all the issues disappeared. New users had no trouble describing themselves in 250 characters, were actually providing us with richer data, and our user acquisition funnel improved significantly.
Assumption 2: Product experience can be primarily over email
We wanted to introduce people to the email experience right away. Although email is ancient, products rarely use it as an interface, making it unique and differentiated. We naively decided we could make the entire experience over email. It’s one of those decisions you look back at and think, “duh, that was stupid”. Email was good for making introductions, but terrible for collecting profile data. People are used to filling out forms. So we gave them what they wanted.
Assumption 3: Deep learning can predict successful intros
There are 7 billion people on earth and 1⁰¹⁹ possible pairs. Humans can’t evaluate all these combinations, and neither could traditional software. We knew the sheer magnitude of choices required a deep learning approach.
We borrowed a deep learning technique commonly used in image classification and repurposed it to a text classification task. At scale, it was really important to get this right, since the quality of introductions would be driven by the accuracy of the model.
Consistently, despite growing our user base, we maintained a 42% acceptance rate to introduction offers. Our model was also showing good performance during training. We attained ~85% accuracy within the domain of technology professionals. This gave us confidence that deep learning was the right approach for predicting who should meet.
Assumption 4: Networking is difficult; people have a networking problem
It’s hard to truly measure this early on, but it’s possible to use proxies. We threw up a webpage with an email sign-up. After some A/B testing and optimization, we were converting traffic at about 15–20%. Even our first landing page, which was just a wall of text like something out of the early internet days was converting at 7%. We concluded that our value proposition was resonating with our target audience.
Lesson: Freemium models may be prone to market traps.
When we dug deeper into our user base and the impressive conversion figures, we discovered a slightly different narrative. We knew there was certainly a market for “introductions” in the startup community, but we were skeptical of the broader market.
In startups, people rely on their networks because networks are critically important for talent, sales, fundraising, etc. However, larger corporations have internal resources that facilitate those needs, and therefore employees are less reliant on their network.
We also think we were experiencing the “early adopter effect”. People signed up because they were intrigued by the technology, but not necessarily the value proposition. Something like, “might as well give it a try; it’s free!”. The freemium model along with high conversions on the landing page trapped us into a false sense of the market opportunity.
Assumption 5: Some people are willing to pay for this service
Perhaps. We had a few ideas on how to monetize, but we needed critical mass before even thinking about experimenting with them. There just isn’t enough value in a network-based business when it’s small. The frequency of introductions is the most likely candidate for monetization. If you want more intros, you pay. Anecdotally, some users were willing to pay, but certainly not a material portion of the existing user base.
Lesson: Few people pay for vitamins.
Initially, we assumed Rey would adequately facilitate both transactional (recruiting, sales, fundraising) and organic relationships (general networking). In reality, users seeking transactional relationships were willing to pay the cost of existing mechanisms that were tailored for them ($50/month for Linkedin Premium, $20K for recruiters). With a small network of users, Rey couldn’t directly compete with those alternatives.
For those wanting to expand their network organically, they would use Rey amongst other free tools. In both cases, Rey was a nice to have (a vitamin); not a must-have (an aspirin).
Assumption 6: High-quality introductions drive viral user growth
Networks generally grow using some viral mechanism. Users sign up, get some value, recognize they would get more value if they invite their friends, invite their friends and the cycle repeats. Some pretty incredible growth figures can be achieved this way; Facebook is a prime example.
Unfortunately, Rey didn’t have a viral mechanism intrinsic to the product experience. It didn’t matter if your friends were on Rey, since you would join to meet individuals outside of your social graph. As a result, good introductions did not generate referrals at the necessary exponential rate (although we did see some).
Lesson: networks with no inherent viral loop are harder to scale.
There is a simple reason virality is important early on: when your annual revenue per user (ARPU) is zero, it automatically disqualifies all costly acquisition channels. In addition, only viral loops can produce the exponential user growth necessary for a network.
Often, there are ways to spark virality even if it’s not intrinsically there. One experiment we planned to run was to make the network exclusive and invite-only. We would then incentivize existing users to invite their friends in exchange for more frequent introductions. A similar growth tactic was used by Gmail in the early days.
We even deployed a website that would communicate this approach:
Unfortunately, we ran out of time and never launched the experiment.
The current state of affairs
Rey alpha was launched in September 2017, with the public beta launched in November 2017. As of today, these are some of the metrics we’ve achieved:
- ~1,000 users
- ~ 7% WoW growth on average with zero marketing spend
- ~100 introductions made
- 42% introduction offer acceptance rate
It’s decent progress, but nothing spectacular. Lack of virality, smaller than anticipated market and a vitamin value proposition all likely contributed to some of these less than impressive numbers. However, there may still potential for a business here. Perhaps our biggest challenge was the inability to acquire sufficient funds.
Final lesson: be cognizant of the local funding environment when picking a business model.
A network-based business with a long path to revenue is a difficult proposition to sell to early stage investors. At the end of our pitch, we’d often hear things like, “you guys should go raise in the valley”. In hindsight, we probably should have listened.
There aren’t many network-based businesses in Canada and therefore it’s just not a model local investors are familiar with. After nearly five months of fundraising, just under half of the $450K pre-seed round was committed.
Ok, so what’s next?
In its current form, Rey requires very little maintenance to operate. It runs in the cloud, we haven’t encountered a bug in several weeks, and its server costs are negligible.
We see meaningful introductions facilitated everyday, particularly in Toronto and San Francisco. And a healthy amount of new users are signing up on a weekly basis. As a result, we are going to maintain it as a side-project for now.
Of course, in the off chance someone is interested in taking it over, we’d certainly be happy to discuss. Shoot me a message!
