In March 2018 I came back from the south of France and joined a startup called PriceHubble. The team was five people in Paris, all of them sharp and excited. I was doing data collection and data engineering. We were scraping the web for things like polygons of zip codes and cities, populations at different administrative levels, and then reinforcing the open data with governmental and statistical sources. We compiled all of that into models that tried to answer one question: what is the right square meter price of an apartment, given an address and a few characteristics?
It was a beautiful company to work for. The founder, Stefan Heitmann, had already exited MoneyPark a few years before, and could fund the early days himself. That is a huge booster in the VC world. The fundraising rounds came quickly after. My colleagues tried a lot of different model families and a lot of different ways to segment 2D space. I won’t go too deep into what is in production today, but it is some kind of regression at the core.
Why I didn’t take another job
In the meantime my own interest was shifting. I had been writing software since the early 2010s as a student, and professionally since 2013. I still loved engineering, but I had this big business curiosity I couldn’t get rid of. So I went to INSEAD for a year. The MBA was transformative.
After it ended, I had a decision to make. I was afraid that if I took a paid, salaried job, I would get too comfortable and never get close to what I actually wanted, which was to be a startup founder. So I just refused to apply for one. I started doing a bit of indie work instead. My first idea was basically what PriceHubble was doing, but for cars: find the real price of a car based on data. I tried for a while and realized it wasn’t going to work.
Meeting Youssef
I went on the Y Combinator co-founder matching site, met a bunch of people, and ended up clicking with Youssef. Together we started RETCO.
The first idea we had was process improvement for people working in commercial real estate. We thought we could help them move faster and stay more on top of their pipeline. Whenever we showed it to a customer, the response was something like “this is useless, I can already do this with an email.” Looking back, they were probably right. They wouldn’t have gotten as much value from our tool as we were hoping.
The pivot: institutional knowledge
We started paying more attention to what these companies were actually losing. We noticed that whenever an employee left, they walked out with a huge amount of context that the company had never written down. Years of intelligence about deals, buildings, contacts, neighborhoods. Just gone.
So we tried to help them keep that knowledge. We had a few successes. Some companies signed. We hit a bit of MRR. Not huge, but not negligible either.
To do this we built what was probably one of the first generations of AI document structuring. We used the early LLMs, starting with GPT-3.5, to turn unstructured PDFs into a structured database. The use case we focused on was opportunities. Investors in commercial real estate receive offering memorandums for buildings that are not publicly advertised. They skim them in a few minutes and dismiss most of them. All of that intelligence was being thrown away, and we captured it.
The problem was that the market just wasn’t ready to spend money on this. We were convinced we could change that. We were wrong.
The Soneka deal
The software itself had real value, and we ended up reaching a deal with Soneka, another SaaS company in the space. They had a few million in ARR, so they were a step bigger than us, and they were able to make us an offer. What Soneka wanted was to take what we had built and apply it to other data structuring problems they had on their side. I can’t say too much about what exactly we are working on now, but we are moving fast, with real industry leaders in real estate, on a product the customer actively wants. That last part is what is most different from the startup experience.
What I’d keep from all of this
There is a lot more to tell. The day-to-day of running RETCO on a tiny budget. Always being one intern away from being stuck. The Station F allies who got us in and gave us all kinds of perks and help that I will never be able to fully repay.
The thing I want to keep from those years is this. The moments that mattered most were the ones where I was alone at home, working. We tend to overestimate the importance of meeting a lot of people and trying to raise money. So much is possible today with very little, and the AI tooling that has shown up over the last couple of years is only making that more obvious.