How "Tinder for Jobs" Became Worth Millions
Part one of a new series called "Lessons from Startups"
Five months after I joined my first startup as a data scientist - I was partying in Vegas with the whole company celebrating our acquisition.
While I would like to believe that hiring me propelled us to success…it’s pretty clear there are other factors at play.
What separates successful acquisitions from failures? The co-founders went from idea to acquisition in 2 years.
And so on one of my own business problems - I wanted to revisit how I stumbled into startups and the lessons I’ve taken away from each one.
One essay stretched into many pages. And now I’m releasing it as a multi-part series. Here is part one - for the first startup I was a part of.
What was Jobr?
Jobr was a “Tinder for jobs” mobile app where job seekers could swipe right to apply to a job and swipe left to pass. The thesis from the co-founders was that in 2014 - smartphone and mobile growth was growing at an astronomical rate and millennials were more likely to apply to jobs through their phone.
So in the very beginning - Jobr first started out targeting high-end tech jobs. Then realized pretty quickly the actual demand for jobs and staffing was in high churn + high ticket temp jobs (think nurses and truck-drivers, etc…). Jobr would partner with third party aggregators like Ziprecruiter or Careerbuilder to get paid on a per job application basis. These third party aggregators would pay us 50 cents to 10 dollars (truck drivers am-i-right) per application and ideally then be able to charge a higher amount on a per posting to companies.
Now - ostensibly you might think this shift to mobile made it a great business. For one - a user forced to swipe left or right on every single job to see the next one created a dynamic where we got a lot of applies. And on every single right swipe - Jobr got paid! So if a user swipe bombed and applied to every job they received, we would make a lot of money.
This however created a long term problem in which after a period of low value applications to our partners, they would get pissed and cut us off from their job feeds. So that’s where I came in, a junior know-nothing data scientist to make sure quality was in place when users applied to jobs.
First Lesson: Startups Make Money
At my previous role - I was stuck in an entry level analytics job at a mid-stage startup where the founders were mostly absent and my boss quit two weeks after I started. And so in lieu of having any work to do - I had downloaded Jobr and unwittingly applied for a data science position they dogfooded on their app.
My onsite interview experience was the first shock to my system. The CEO had invited me for a “quick tour of the office” and on a cold Tuesday evening I walked into a small studio / loft with 6 dudes sitting together with headphones on locked in to their screens.
The garage myth was real. Next to the open space there was a ping pong table, small conference room with no ventilation, a staircase that led to a loft with a dirty old mattress, and a small kitchen with beer in the fridge. The quick tour of the office turned into a 4 hour impromptu technical onsite interview. And a few months later, they hired me after one of their existing engineers suddenly quit on the spot (before he vested too).
Every day we would get in at 9am and work until 6 or 7pm every night. After the first few weeks, I was dying for more meetings and social interaction. But this is how I learned my first lesson about startups: “We are here to make money”, the CEO conveyed. There was no free lunch and no massages like my last company. And there was also no lack of work to do after I started.
Second Lesson: Impact ≠ Effort
When I started - the CTO had build the initial V1 of the job recommendation engine to get the app up and running. It matched users to jobs by classifying every users resume and job into different industry buckets (think software, education, construction, etc…).
This classifier worked but we were getting relevancy complaints. For example, if my resume said something like “Data Engineer” - I would be classified into the software / tech bucket and shown only jobs in tech. But this meant it would equate a product management or design job to one in data engineering as well.
So to solve that - we indexed all of the jobs into elastic search and built an advanced search query using the user’s position history to sort and return all the most relevant jobs first. Immediately relevancy got better and more accurate for each persons resume. But now we were dealing with another issue. Many users were now running out of jobs if they didn’t have enough experience on their resume.
What solved this problem was probably the most non-data science thing I thought of.
“Why don’t we just take every word in the person’s resume and put it into elastic search?” I brainstormed.
Turns out - adding more keywords from the user’s resume into the search query would inevitably return a higher recall of jobs. Immediately we saw a 10% increase in job applies per user.
The funniest part was that I believe this was my first “real win” and I didn’t think anything about it. I didn’t know the quantification of my efforts until my mentor at the time conducted a yearly performance review.
“By increasing jobs per applied by 10% - you essentially helped us increase revenue by 10%”, he said.
“Huh - but that was the easiest project?” I thought in my head. Turns out this is almost always a recurring theme.
Hard ≠ High Impact
At some point - there was a consensus that since I was a data scientist and since we had millions of datapoints of right and left swipes, that it was time to build a deep learning algorithm that would more accurately recommend a job for any variation of resume.
But turns out a new grad building a deep learning classifier from scratch was a very VERY hard task, especially at a startup with only four engineers including the CTO and myself. Even when I did manage to build the infrastructure, use all 16 CPUs on my a laptop to train the classifier, find some way to host the model on an AWS instance - we would run the A/B test on our new recommender system and it still only saw marginal improvements.
The reason why? Largely performance issues. Surprise, turns out I didn’t actually know how to build good ML infrastructure. In large cities like the San Francisco Bay Area, with a job radius of 25 miles, we were classifying the user’s resume against up to 100K+ jobs and then sorting them by relevance. This could take upwards of a few seconds on just the classification, which then would double our latency for the system which destroyed our user loading times.
So in 2016, what I realized pretty quickly was that if I wanted to do data science at a startup, I would have to first learn ML infrastructure. But I hated ML infrastructure, and no one had time or expertise to work on it. So we abandoned that project after a month of testing and work with nothing to show for it.
There’s Low Hanging Fruit Everywhere
I was still itching to build a classifier that would work and one day stumbled upon an idea. We had a “concierge chat bot” that our growth hacker would manually impersonate to send messages to users. Over lunch he told me that 90% of the user’s questions fell into 5 different categories where he had a complex keyboard shortcut system he would use to copy and paste different answers for various user questions as fast as possible. When the users got their questions answered, they were usually pretty happy, and then he would immediately ask them to “Give us a 5 star review on the app store for good job karma”. Somehow this would always work to boost our rankings on the apps tore.
So over the next few weeks, I shipped a pretty easy supervised classifier that would classify our users questions if they fell into our a FAQ of five or six questions. Afterwards, if we detected positive sentiment from solving their inquiry - we would immediately ask for five stars. The growth hacker stopped manually responding to candidates, and we probably got thousands of 5 star reviews from this one feature that kept our app high in the Apple app store.
I think what I realized that similarly to that classifier - for me 90% of my achievements at the company came from the elasticsearch keyword stuffing and this chatbot. The rest wouldn’t pan out exactly as required. Or it meant fixing technical debt - then it was just in general harder to quantify. But what it showed me was that only a few projects can return the value of a full time engineer.
Third Lesson: Finding ONE Piece of Alpha
In June of 2016 - Jobr got acquired by Monster for $12.5 million dollars. Even better, the executives consciously paid for the team of 10 guys and one woman (plus her bf) to party our brains out in Vegas (with table service).
But then 8 years later - and somehow I’m seeing many companies continue to raise money off of the same “Tinder for Jobs” concept.
So what happened? Why are these guys reinventing the wheel? Once Uber gets built, no one else can enter in right?
Well….maybe not. You could say that the job market is quite big and they aren’t winner take all markets. You could also theorize that as app development gets increasingly easier - these “tech startups” become more or less commoditized apps that anyone can build. But more importantly I believe they were probably exploiting the same strategy we used to grow Jobr and get acquired.
When Jobr first got started in 2014 - it was peak mobile Facebook ads if you could crack it. And realistically, Jobr got acquired through our efficiency in acquiring customers through Facebook Ads. Monster’s executives had just botched their mobile app build and launch. And so one day while scrolling on the app store one day found Jobr at the very top of the app store somehow under “job search / industry”.
We managed to snag this position in early 2016, and for the one week after Facebook launched Instagram ads, we were getting extremely low cost per downloads. Like stupidly low. So our ad agency guy started juicing the spend!
Here were the unit economics → on the best month of the year Jobr paid $1.08/download. For the Monster executives their best month was $10.76/download.
So in some sense - it was not off of any kind of new technology as much as the fact that we were a customer acquisition machine that Monster couldn’t replicate and figure out on their own. And my sense is that these new job search apps are riding off of a similar wave on TikTok and exploiting a very simple strategy of finding novel ways to acquire customers.
There were probably some other reasons. Yes Monster probably did want us to replace their terrible mobile app. But at the end of the day it was a guy working part-time remotely in Thailand who probably had 10+ other Facebook paid ads clients that ended up really helping us get acquired.
Naturally he also joined us on the Vegas trip.
Last Lesson: How acquisitions can destroy the existing team
After the acquisition, it eventually took two years for Monster to fully integrate us into their company. But only one year for the existing team to actually fully check out.
This was the conundrum of any kind of company acquisition. You must align the incentives between the parent company and the acquired company for as long as possible.
But succession is always hard. And sometimes there’s easy ways to pitch the existing team members to stay on.
One year after joining Monster, the Jobr CEO left, but before leaving pitched an idea to the new Monster private equity boss. He postulated that his retention bonus that would be forgone should be distributed to the rest of the team. It took three months to really negotiate it to almost going through. But at the last moment, the new incoming CTO of Monster nixed the deal to save budgetary costs. And from then on out - the existing team was basically a shell of itself.
Put yourself in my shoes. My old manager drunk at a bar had told me that my total comp at age 23 was going to be almost $200K+ when the deal was almost signed. And to hear it was then actually not going to happen a few weeks later destroyed my motivation.
I remember the last straw came one day when random workers came in to pack up all the snacks in our office. Their actions were going for the jugular of startup culture. And from that moment onwards I was checked out.
I barely did anything at work. I would join standup and make up projects that I was working on. I spent as much time as I could working from home and taking calls from the beach as I improved my surfing. I even got close to becoming “over-employed” and taking a second job from LA but ultimately failed the onsite interview.
It’s almost always a matter of time thing for big companies to erode the benefits of startups. But mostly because the incentives are almost never right for private equity style acquisitions. If they had just paid me the typical FAANG tech salary I would have stayed and probably been more slightly more motivated to continue our bet of increasing market share.
Money truly is a strong motivator when you’re paying above market. And if you never build a startup - you don’t get a chance to really understand what that value is.