Tractable applies AI to accident and disaster assessment – TechCrunch

“Happy to spend 10 minutes on our vision and the journey we’re on, but then, really, 15 minutes on what we have today, what we’ve accomplished, what our AI is doing, ”says Docile co-founder and CEO Alexandre Dalyac when I called him a few weeks ago by video. “You can probably speed it all up,” I replied.

The resulting conversation, which lasted well over an hour, covered all of the above and more, including what is needed to build a successful AI business and why he and his team think they are. can help prevent another ‘AI winter’.

Founded in 2014 by Dalyac, Adrien Cohen and Razvan Ranca after going through the business builder Entrepreneur first, Tractable, based in London, applies artificial intelligence to disaster and accident recovery. Specifically, through the use of deep learning to automate the assessment of visual damage, and therefore help speed up insurance payments and access to other types of financial assistance.

Dalyac is embarking on what is clearly well-prepared and obviously polished ground. “We are on our way to helping the world recover from accidents and disasters faster. Our belief is that when accidents and disasters happen, the response could be 10 times faster thanks to AI. So what we mean by that is everything from road accidents, burst pipes to large-scale floods and hurricanes. Every time any of these things happen, things get damaged. ”

These things, he says, break down globally into cars, homes, and crops, which equates to about $ 1 trillion in damage every year. But, perhaps more importantly, livelihoods are affected.

“If a car is damaged, mobility is reduced. If a house is damaged, the shelter is reduced. And if crops are damaged, food is reduced. Through all of these accidents and disasters, we are talking about hundreds of millions of lives affected. “

This is where a little lateral (not artificial) thinking is needed. Crash and disaster recovery begins with a visual damage assessment: look at the damage, say how much it will cost, unlock the funds, and rebuild. The problem (and Tractable’s desirability) is that for an assessor to examine a car, house, or land can take days to weeks depending on availability – and therefore can access funds to begin rebuilding – while the claim is that computer vision and AI technology can potentially do the same job in minutes.

“When you assess, it’s basically a very powerful but very narrow visual task, that is, look at the damage, how much is it going to cost? Today, as you can imagine, this type of assessment is manual. And it takes days, even weeks. And you know that instantly with the AI ​​which can be 10 times faster, ”says Dalyac.

“In a certain sense, this is a perfect class of AI tasks, because it is very heavy on the classification of images. And image classification is a task where AI can surpass human performance from this decade on. If you have an instant assessment, it means faster recovery. Hence the mission.

Dalyac says some of Tractable’s secret sauce is found in the millions of exclusive labels the company has produced. This has been facilitated by its patented “interactive machine learning technology”, which allows it to label images faster and cheaper than traditional labeling services.

To date, the team’s goal has been to train their AI to understand damage to cars, a technology they have already deployed in six countries, with the startup working primarily with insurers.

Related to this, I am shown a simple demonstration of Tractable’s car damage assessment tool. Dalyac opens a folder of car images on his laptop and downloads them into the software. Within seconds, the AI ​​apparently identified the different parts of the car and determined which parts can be repaired and which parts need to be fully amortized and therefore replaced entirely. Each has an estimated cost generated by the AI.

It all happens in a matter of minutes, although I have no way of knowing how difficult the predetermined and fully controlled task is. It’s also unclear how an AI can possibly do the full job of a human evaluator based only on a limited set of 2D images, and without the ability to peek under the hood or undertake further investigations.

“We try to determine the damage to a vehicle based on photos,” says Dalyac. “There are some really difficult correlations to identify, namely: based on the photos of the exterior, what is the internal damage? When you’re a human you’re going to have seen and wrecked about a thousand to two thousand cars in your 20 or 30 year life doing that. Our AI has already been trained on tens of millions of these cases, so this is a perfect case where we have already distilled the work experience of thousands of people. This allows us to get our hands on some very difficult correlations that humans just can’t make.

That said, he admits that a photo doesn’t always contain all the information it needs and can only have a certain level of precision. “You may then need to have the car dismantled and photos of the internal damage taken. You might even want to get some data from the dashboard. And you may think that as cars have more sensors… the assessment will not only be visual, but also based on IoT data. But that doesn’t take away from the fact that we’re confident that AI will do all of this. ”

What is perfectly clear is Dalyac’s commitment to developing AI technology with real-world use that is commercially viable. If that doesn’t happen, he thinks it won’t just be Tractable that will suffer, but the continued belief and investment in AI as a whole. Here, of course, he talks about the prospect of another so-called ‘AI winter’, citing a recent Crunchbase report This indicates that funding for artificial intelligence companies in the United States has stabilized and even started to decline at the seed stage.

“If you’re trying to make sure that the $ 15 billion that has been invested in AI doesn’t screw up and lead to something successful that will prevent an AI winter that will lead to continuous improvement, you have need a very good return on this asset class. And for that, you need these businesses to be successful.

“For an AI business to be a success, a real success – not just an acquisition, not just an IP release, but a real business success that will prevent an AI winter – you have to find cases of concrete uses that will make the difference. , where you can surpass human performance, where you can change the way things work, ”he says.

The reference to IP acquisition or output makes more sense when you consider that Tractable was in the same cohort at Entrepreneur First as Magic Pony Technology, the AI ​​startup acquired by Twitter for up to $ 150 million for its image enhancement technology. And more recently, the team behind Bloomsbury AI, another EF company, was acquired by Facebook for $ 20-30 million.

To ensure Tractable can continue its mission of applying AI to disaster recovery and disaster recovery – and presumably not to sell too soon – the startup closed a $ 20 million Series B investment in a cycle led by the American venture capital firm Insight Venture Partners. Existing investors including Ignition Partners, Zetta Venture Partners, Acequia Capital and Plug and Play Ventures also participated. New capital must be spent to accelerate growth, develop research and development and enter new markets.

(Series B also included additional secondary financing of $ 5 million, which allowed some investors to at least partially opt out. I understand that the founders of Tractable sold a relatively small number of shares as they were allowed to withdraw money from the table. Dalyac declined to comment.)

As we wrap up our appeal, I note that all of Tractable’s major investors, except EF, are from the United States – which Dalyac says is a deliberate move after discovering the divide between European and US valuations. .

“It’s a shame, isn’t it?” I say with my European technological ecosystem hat.

“This is not the case; these are huge exports to the UK, “says the CEO of Tractable, born in France but raised in the UK.” We currently have the vast majority of our workforce in London. The entire product team is in London. The entire R&D team is in London. But most of the income comes from the United States. We are making AI an export industry from the UK ”