How to build an AI unicorn in 6 years
How I did it (and you can do it) starting with zero knowledge of tech
Today, Tractable is worth $1 billion. Our AI is used by millions of people across America, Asia and Europe to recover faster from road accidents. It helps recycle as many cars as Tesla put on the road in 2019. And yet 6 years ago, Tractable was just me and Raz, two college grads coding in a London basement. A year before that I knew nothing about tech. If it’s happened to me, it can happen to others, so here’s the story & learnings.
Part 1: From AI rookie to techie
In 2013 I was finishing a college degree and starting a company was the dream. I’d tried building a fashion crowdfunding startup with friends, so independent fashion designers could get their designs funded and sold online. It had gone nowhere.
That’s when I heard about Entrepreneur First (EF). Now a world-leading talent investor, EF had then just kicked off as a company building program for ambitious college grads. Perfect. I applied, talked about fashion crowdfunding and got rejected — because I “knew nothing about tech”. EF was a tech company builder, because as Matt their CEO put it: “in the past, if you wanted to build a product for millions of people, like a car, you needed an entire factory. But now with computers and the internet, one person in their bedroom can write code to serve millions of people globally. This level of scalability is unprecedented”.
I’d seen software as dry, rigid and frustratingly intangible. Now it was company building magic. Finding that Imperial College offered a computer science conversion course, 2 months later I was enrolled for a September start. This surprised the housemates. “It says modules in ‘data structures’, ‘relational databases’ and ‘compilers’ — you don’t even understand what these words mean. Are you sure about this?”
Looking back, that year was harder than starting Tractable. But that Imperial degree changed my life, so if you’re an aspiring tech founder CEO I bet it could change yours too.
It first took me weeks to write a program in C++ that simulated a chess game — just printing text saying where the chess pieces were on the board after a move. I remember a friend saying “isn’t this chess program kind of dumb? are you sure you’re going to be able to build a tech company after this?”.
I started (learning Python and) taking a course on coursera¹ called machine learning with neural networks by Geoffrey Hinton, the father of deep learning². Honestly, it was like being love struck. Back then, to me AI was just science fiction from Terminator. And yet a Wired article said the field was resurging, via multi-layered artificial neural networks, aka deep learning. Wow. And it turned out to all be based on statistical regressions (linear algebra, calculus and statistics) — just like what I’d studied in mathematical economics*, only with a million times more variables and data. I couldn’t believe that you could teach a computer to see by using the same maths as what economists use for predicting things like employment. It’s still a wonder to this day.
In the 2nd term, teaming up with much more skilled people, we built a plant recognition app with deep learning³. I’ll always remember us walking Prof William Knottenbelt to Hyde Park, seeing him take photos of flowers with the app and laughing from joy as the AI just worked. It was recognizing the right plant species. This had previously been impossible, especially in the ‘cluttered background’ setting of a park where all kinds of other objects, shadows and lighting in the image would throw off traditional computer vision algorithms.
I started spending every sparable moment on image classification with deep learning. Still no one was talking about it in the news, even Imperial’s computer vision lab wasn’t yet on it! I felt in on a revolutionary secret. Over the summer I got deeper into research — under the supervision of Prof Murray Shanahan and Dr Jack Kelly — working on image classification for water & gas utility inspection, as well as certain special properties of neural networks, which culminated in a PhD offer.
Looking back, narrowly focusing on a branch of applied science undergoing a breakthrough paradigm shift — that hasn’t yet reached the business world — changed everything. If you want to be a deep tech entrepreneur, I’d do that.
Part 2: From AI techie to founder
This time, knowing something about tech, the application for EF worked. The last interview at the time was a hackathon, and that’s where I met Raz. He was doing machine learning research at Cambridge with Prof Zoubin Ghahramani, had topped EF’s technical test, published papers on poker bots that could detect bluff, and on reconstructing shredded documents. His bare bones webpage read: “I seek data-driven solutions to currently intractable problems”. So cool. (That’s where we’d get the name for Tractable). That hackathon, we coded all night. The morning after, he & I knew something special was happening between us. We moved in together and would spend years side by side 24/7: from waking up to the heavy metal tunes of Pantera in the morning to marathons of coding at night.
We joined EF with a headstart. I knew our mission was to bring the image classification breakthrough to the real world. And thanks to an intro from Jack Kelly, we had a first paying customer. Our first use case was… plastic pipe welds. It was as glamorous as it sounds. Pipes that carry water and natural gas to your home are made of plastic. They’re connected by welds (melt the two plastic ends, connect them, let them cool down and solidify again as one). Image classification AI could visually check people’s weld setups to ensure good quality. Not glamorous, but real world value for breakthrough AI. In the end, they — our only paying customer — stopped working with us in the middle of our first fundraise. That was rough.
And yet somehow $1.9M was raised. It started with Matt saying one day: “someone’s coming in to visit, a Google early investor. Win him over and you’ll be set”. That person was Charlie Songhurst. As a 20 year-something consultant he’d advised a tech giant to buy Google; the client didn’t, but he put all his savings on Google call options shortly before the IPO. Let it be said that he didn’t really need a job anymore after that and has been angel investing full time ever since. Hearing our pitch, it seemed Charlie didn’t care for the tech and business details so much as our personalities; what got him excited was that Raz & I liked to free-climb ie take calculated risk. He started off with a $50k commitment, but the more American angel friends he introduced, the easier it became to win a $100k+ commitment from the next. One of them called on the phone for 15 minutes, and committed $300k. To this day I have never met this person.
An “angel party round” was great, but Matt thought we needed the disciplined guidance of a venture capital (VC) fund. That would be Ash Fontana, from Zetta Venture Partners in San Francisco. He’d seen our pitch video, done calls with us but did not feel ready to make a move After speaking to a couple insurance prospects and hearing that we were about to sign the party round, he flew from San Francisco to London the next day and wouldn’t leave until we’d let him in to lead the round⁴. Once you’re ready to close your fundraise, fear of missing out will unlock more offers from great investors. On the big day, we were in Berlin staying on a boat hostel with no means to print & sign the investment docs with a witness. We had to find a print shop and ask people on the street “hello can you write your signature here so we can get $2M for our company”, which for some reason took quite a few attempts to succeed.
Part 3: Finding product-market fit
300k pipe welds inspections a year in the UK was too small a market, so we explored image classification use cases in utilities, geology, dermatology and medical imaging, and landed on car insurance. There were not thousands but billions of car damage images produced a year by repairers and insurers: an AI training data goldmine. What if after a car accident, people could take photos of the damage with their phone, and let the AI handle their insurance claim automatically, without hassle? It took us a full seed round to sign our first partnership with Mitchell, a US software company⁵. We’d trained AI to recognize a car’s front left wing (aka fender) on a photo, and say whether it needed to be repaired or replaced. The front left wing would grow to become our mascot, the gift of choice to honor team members for services rendered.
Raising our next round, an $8M series A, was hardest⁶. You normally need one million dollars of recurring revenue by then; we were far from that, and cash was running low. Winning customers in insurance was slow; we weren’t making the numbers despite the Mitchell partnership and a pilot with Ageas, a UK insurer⁷. There were so many vulnerabilities in our pitch that we had to refine the words (and tone of voice) used to answer specific questions over what felt like a hundred iterations. Thankfully, our seed investor Ash Fontana’s hustle was there to help, making intros and following up with prospective investors to create additional momentum.
It wouldn’t have happened without Adrien, our 3rd Tractable cofounder, who’d joined right after seed round. Adrien had previously cofounded Lazada, an online supermarket in South East Asia like Amazon and Alibaba, which sold to Alibaba for $1.5B. Adrien would come to teach us how to build a business, focus, inspire trust and hire world class talent. We’d met 3 years prior, when I’d interned at Lazada; I’ll always remember sitting next to him at a team dinner in Ho Chi Minh City and bonding around his stories of photography and acting. I had got pretty emotional on the last day; he’d become an older brother figure. Him joining as Tractable 3rd founder has probably been our greatest blessing.
And yet despite the series A, we didn’t know if we had a business. It had taken twelve months to sign a one month pilot with an insurer and barely covered the travel costs incurred. We had tried to get acquired for $25M by a tech giant (which looking back, thankfully didn’t work out). Around that time, soon after a proof-of-concept with an insurer gone wrong, one of our best AI researchers quit.
Signing the first million dollar customer contract changed everything. It’s rare for a startup to sign that size, so that’s when we really started believing again. By showing an insurer that our AI could generate $50 worth of speed and expense reduction on each of their 1 million claims, a $1M contract to get going was well worth it. By then, AI had become an exciting technology for the public, and Mitchell had made us credible. A year ago I had ducked under the table to stifle nervous laughter as Adrien asked a prospect for $30k/month, now I was bringing back $1 million of business to feed the Tractable family for months.
Doing that 3 times over brought us into legendary “10x year on year growth”. When time came to raise the series B round, compared to series A this roadshow was like riding a warm knife through butter, culminating in us raising from the almighty Insight Partners⁸. They too were introduced by Ash Fontana.
Part 4: Scaling & positive impact
Scaling to a $1B valuation company has been fueled by winning a large contract with a country’s leading insurer and announcing publicly. Many large companies will hesitate to adopt a disruptive solution offered by a newcomer, even if your product is best performing. But if you can succeed with one of them and agree to a public announcement, that will be the seismic shift: now all of their competitors will risk falling behind. Now many will be thinking about adopting your solution as a fast follower. After discovering this in Japan, we went on to repeat this approach in France, Poland and most importantly the US.
Expanding to Japan was quite something. I will always remember the first dinner in Chicago with Hidenori Kobayashi from Tokio Marine, Japan’s oldest insurance company. Our intermediary Ted⁹ suddenly had to be pulled away, leaving us two at the table, lost without translation. We managed to make ourselves understood, and although words were scarce we could see we had similar ideas. Was it his first long 1:1 in English? It created a bond I hope to keep. Hidenori and his team¹⁰ were the only customer to regularly fly to London to work with us from our HQ office. Once they gifted us a Daruma: a doll symbolizing perseverance and good luck. We colored the first eye to mark the start of our partnership, but the 2nd eye could only be colored once our goal was met. 18 months later, after going live with our AI without any human intervention, and being accepted by Tokio Marine’s shokunin appraisers, the 2nd eye was colored.
We were proud to be building a fast growth business centered on cutting edge AI, yet the team kept asking what our mission was. We’d say “to build a real world AI business”¹¹, but as a team member pointed out over drinks, it didn’t help them impress the opposite gender. Even if that sounded a bit coarse, it was downright honest and powerful. It taught us that a great mission needs to be inspiring, not just personally but yes, also in social settings. Getting that right meant attracting and retaining the best talent.
We started paying attention not just to growth, technology and value creation, but also positive impact. We realized that when crashed cars were too expensive to repair, they were sold for scrap at online auctions, and our AI could help figure out which ones to recycle¹². Every car recycled for parts is about half a metric tonne of carbon dioxide emissions avoided¹³. Having our AI analyze photos of damaged cars on auction — and help suggest which parts are in good enough condition to be recycled — has materially increased the economic value of recycling cars for LKQ, the world’s largest auto recycler.
We also realized that climate change creates natural disasters like typhoons and hurricanes, wrecking homes and generating many insurance claims that our AI could accelerate¹⁴. In our earliest days, a brilliant intern used free compute credits and idle servers to mine ethereum cryptocurrency during its first days. After holding it for 6 years, we’re selling the ethereum for millions of dollars to create an AI disaster recovery fund together with our investors at Georgian¹⁵. It’ll allow us to deploy our natural disaster appraisal AI to those who need it most, regardless of whether they can pay.
There continued to be really hard moments, where initiatives fail painfully, or when a Board member spends a year trying to get you to step down. But what doesn’t kill you makes you stronger.
Some of the AI out there can be marketing tag lines without substance, systems that don’t yet work well, or technology with controversial impact on privacy. We develop AI that over time works for real, respects privacy, and is hopefully reducing hassle for millions of people. AI will transform humanity this century, and hopefully Tractable will play a meaningful role — for good.
—
If you want to learn how to become an entrepreneur and want some practice before EF, join us at Tractable! Lots of us are entrepreneurs-in-residence who one day will start their own startup.
Follow me on LinkedIn and Twitter.
Too many people to thank who made this possible, but to name a few: Matt Clifford (& EF team), Andrew Ng, Imperial Computing (Jack Kelly, Will Knottenbelt, Murray Shanahan, Whatplant team), Zoubin Ghahramani, Tractable (Razvan, Adrien, Julie, Ahmed, Nat McAleese), Investors (Charlie Songhurst, Ash Fontana & Zetta team, Insight team, Ignition team, Georgian team), Mitchell team (Beau Sullivan, Scott Kozak), Rob Smale, Hidenori Kobayashi & Tokio Marine team, LKQ team (Yogi Shivdasani, Terry Fortner), and of course… all our customers and the entire Tractable team.
—
Footnotes:
¹: Thanks to Andrew Ng’s coursera, anyone can learn from a world expert.
²: Jack Kelly’s encouragement gave the confidence to try what seemed hard.
³: Imperial College got its students to build truly industry-grade products.
⁴: Thank you Ash Fontana, Mark Gorenberg, Jocelyn Goldfein for your support.
⁵: Thank you Beau Sullivan, Scott Kozak for bringing our companies together.
⁶: Thank you Nick Triantos, Scott Coleman, John Connors for your support.
⁷: We owe it to the visionary leadership of Rob Smale, Ageas’s head of claims.
⁸: Thank you Teddie Wardi, Lonne Jaffe, Jeff Horing for backing us early.
⁹: Thank you Ted Ohkuma for bringing our companies together for the first time.
¹⁰: Thank you Masaki Ishihara and Yousuke Oohashi for your dedication & trust.
¹¹: We believe in building an AI business to help avoid the next ‘AI winter’.
¹²: Yogi Shivdasani and Julie Kheyfets, I will never forget coming up with this together, as a result of Terry Fortner’s introduction.
¹³: Auto recycling environmental impact from research by Waseda university.
¹⁴: Ahmed Hameed, I will never forget coming up with this together.
¹⁵: Emily Walsh & Madalin Mihailescu, thank you for your backing.
