This article was automatically translated. Read the original in french.
“While doing the laundry, I found a sweet note addressed to my wife” While for a human, it is obvious that this sentence speaks of infidelity, for a robot it is far from clear. It is with this challenge, to understand and support people whose relationship is in trouble, that Deeplink was born.
As explained in the first episode, I had decided to build a tool around conversation and artificial intelligence (AI). Now I had to materialize the wish… but where to start?
It may sound a bit crazy, but my first project as an entrepreneur will be… to talk about couples’ problems. This is the challenge that Patrick Delarive and Gérard Decombat offered me: to create a psychological chatbot.
Obviously, the naive enthusiasm of the novice entrepreneur is quickly followed by doubt. How am I going to make a robot understand that humans have a lot of different expressions, especially when it comes to talking about delicate and personal subjects?
- My wife cheats on me
- I am a cuckold
- My husband sleeps with the neighbour
- I have horns, I don’t go under doors anymore
- My children look like the postman
What a hassle! All these expressions speak about the same subject: infidelity, and no body of data exists for this subject. Worse, when I looked for other examples on psychology blogs, I observed that no one displays their feelings as raw as when they talk to a shrink, let alone a robot. Take voice assistants like Alexa or Siri: who has never asked a naughty question, or insulted the system? Facing a robot, we express ourselves without filters.
Our first good idea: probe
Daniel Peppicelli joins the adventure, and we probe our entourage: we discover some expressions, but it cruelly lacks diversity. We need other real people; we have thirteen topics to cover such as sex, jealousy, the future, children, separation, violence… That’s when we have our first good idea: we create the 1st data collection tool for chatbot in the world. Through small contextualized surveys, we collect a collection of real sentences and expressions, from humans talking about their relationship: the ideal way to train our AI. The success is there: in three days, we collect more than 5’000 sentences!
Preparing AI is good, but Gérard, our psychologist, also needs to transpose all his knowledge of the field into a program. We test some tools on the market but it’s a cold shower: complex, obscure, requiring to know how to code, they are not adapted to a non-computer scientist. We then decided to use a visual editor, in other words a “no-code” approach. There is no more technological brake, Gerard writes more than 1’800 answers in 6 months, which will give birth to the launching of Psykoople which has since responded to more than 45’000 interactions in French alone.
On the Deeplink side, these two first features will become the basis of our product. Then Covid arrived. We proposed to the canton of Vaud and the canton of Geneva to set up a chatbot informing about health restrictions and aid granted. They accepted enthusiastically: thanks to our tool, in 24 and 72 hours respectively, the chatbots are online. In the first few days, we observe up to 500 interactions per hour. There is no doubt that our tool is really useful.
So the machine was launched, it was time to accelerate and think about the first fundraising. But as you will discover in the next episode, not everything went as planned..