“Going from an enterprise to a start-up was a fascinating experience. The most dramatic difference is that there is nowhere to hide. In the enterprise world, there is always someone to do something for you. You can flag a problem and pass the responsibility for solving it to a specialist — you know that somewhere down the tree it will get resolved. In the start-up world, there is no fall back — you go and do it yourself. Your knowledge is constantly on display. It is important to openly admit where your knowledge starts and stops because if you try and hide, or paper over gaps, you will quickly lose credibility. I think it is an exhilarating experience being this close to the bleeding edge of production.”
Robin: How have you, and the Airtasker team, approached the level of growth the company is experiencing?
Paul: Start-ups generally start in survival mode — just getting things out there. Eventually, you need prepare for growth and transition to a more sustainable position. There are really three categories you have to focus on when preparing for that kind of continual change — people, processes and the platform. Everything comes down to constant expansion. We had 15 new starters last month. It is important to have a robust onboarding process and a constant pipeline of talent. From the technical side, your platform needs to have the capability to expand and handle the growth in volumes. My role has been to develop the processes, technical-stack, architecture and roadmaps necessary to make sure that the team is able to run and grow as fast as our consumer base expands.
The biggest thing we are working on that doesn’t fit traditional approaches to high-growth is around machine learning. Our business has created a need to categorise and review every task that is submitted. At the moment, this is a loose process with an extensive manual component. We have certain categorisation criteria we have to meet for insurance purposes — but we currently do not require categorisation of tasks because we don’t want to limit what can be done on the platform. We want people to be able to use Airtasker to outsource jobs that we haven’t even thought of — everything from end-of-lease removal to queuing up for an iPhone.
Carl and Alan are our long term answers to this problem. They are machine learning algorithms. Carl, named for Carl Linnaeus the father of taxonomy, is our categoriser. Alan, named for Alan Turing of Enigma fame, is our decoder.
Right now, Carl puts everything into 120 category buckets. That helps us with insurance, but also helps with marketing. We want to be able to track what you type on our website while it is happening and advise you on how to best describe what you need. The whole idea is that we want to make sure your task is described as appropriately as possible so that you will have the best chance at finding the right person to complete it. If people can bid accurately, you will get more bids, and more relevant bids — improving the experience from both the poster and worker side of things.
The first phase of Alan is to screen posts for mobile numbers slipped into bypass the system. We have developed some interesting algorithms around that. People will misspell words in a way that still phonetically makes sense. We are using a third party API to do text to speech, and then back to text to pick up the phonetics. That, however, is the easy part. The next phase is marketplace rules. For example, we allow tutoring on the platform, but we wouldn’t let someone offer to write an essay. So, we need to find out ways to discriminate subtle differences across a number of lines in an automated manner.
Robin: How did you go about starting this development process and what is your long-term goal for these programs?
Paul: Our first step was to use PredictionIO, which is part of Salesforce. We just fed training data into that using Upwork to outsource the categorising of about 200,000 tasks by hand. That was fed into the machine learning service. This system got us to about 80% accuracy — but we then hit a wall. It was at that point that we decided we had to make data science a core capability. Our priority in this process was presentation. The communication of data to a layperson is really the crucial factor in being able to act on the information gleaned by the data scientists. Taking this approach, I think, has not only allowed us to build a more useful system, it has allowed us to attract talented data scientists. For them, being able to put their work into practice and actually see that output in their day to day lives is a draw. This focus on creating a user interface that sits on top of the machine algorithm — allowing people to actually play with the data and feel comfortable — feeds into the naming we came up with. In our first sessions, we called the systems something like Air Classification. But, we gravitated towards actual names — Carl and Alan — because we wanted it to ‘become a person.’
My plan is to take Carl and Alan open source. I think there is utility in open collaboration. More importantly, we have benefited a lot from open source software, Scikit, for example — giving back is the right thing to do. It also builds credibility. We are fairly well known in the start-up community — we want to be well known in the technology community. We want to get to the point where when we are searching for talent, people are already sold on the company because we have a reputation for innovation — that has value in as much a real sense as proprietary software.
Robin: In a broader sense, what do you see as the next move for Airtasker?
Paul: For Airtasker, we want to expand overseas at some point.
Most of our campaigns in Australia have been done on TV. This is one area of marketing that has really surprised me — we can see a direct and almost immediate correlation through television advertising and customer growth, often the same day. The interesting thing, and I think that this is related to the utility of our product, is that the growth sticks. We will run an advertising blitz, it will stagnate, but the growth in the customer base will remain. This has allowed us to continually build on the size of the operation. I think we can take this growth model abroad and develop a self-sufficient community through a period of sustained advertising.
I think the question in new markets is if our platform will be conceptually accepted. There is actually a huge cultural aspect to Airtasker. In Australia, we showed up just after Airbnb and Uber gained traction. People were suddenly comfortable hailing a stranger’s car from their phone or staying in someone’s house they met online — all basically based on nothing more than a digital reputation. If you think about it, those are both extremely odd things to do. We were able to capitalise on that consciousness shift regarding who and what can be trusted. It is not a hard sell once people are on board. In reality, you can learn a lot more about a person based on a few dozen online reviews than simply based on the fact someone took out a paid advert in the Yellow Pages — you are also very likely to get a higher quality of service online. However, there is a tipping point that has to occur for people to even consider that possibility.
If you look at it from the larger perspective, Airtasker requires a high-trust culture. Our research indicates that only Canadians outpace Australians when it comes to trust. We believe in the goodness of people and are deeply and profoundly shocked when that is not the case. Our business leverages that cultural perspective. As long as you have some stranger recommend that person, you will trust them. Would this work in South Africa or Mozambique? I don’t know. The three pillars of any market are liquidity, access and trust — we need all three to operate.
Paul Keen is the Chief Technology Officer at Airtasker — an Australian digital outsourcing community that is bringing the convenience of Uber and Airbnb to hiring a handyman or finding a temporary project manager.
Paul joined the start-up world 18 months ago when he moved from Dick Smith to Airtasker. Since, he has overseen the construction of a technology platform capable of sending out 300-400 notifications a second, every second to accommodate for the continual communication-based growth Airtasker has undergone. Since its founding in 2012, Airtasker has expanded to host over 1.5 million people on their platform and continues to experience 25% growth every quarter — entering the category of hypergrowth.
For his latest interview for TechBoard, Robin Block from MitchelLake sat down with Paul to gain insight into how Airtasker is using machine learning algorithms to increase their growth potential and gain Paul’s perspective on start-up culture and the future of Airtasker.