The first time I was deeply drawn into an idea was in my second year at university. I took a module on the theory of networks because its tutorials taught Python.

A network is a set of nodes (points) connected by edges (lines). The edges can have a value and a direction.

That same year, I was also taking a module on economic philosophy, which made you question: What exactly is the economy?

Aha! A is, say, a person who makes shoes, B is a person who makes plants. So A exchanges 4 shoes for 5 plants. The economy is just a (very complex) network!

The course was graded through a research paper on a topic of your choice. I decided to write a Python program that simulates an economy as a network.

The project had a profound impact on me. It fuelled my curiosity, passion and drive towards finding and investigating interesting and beautiful ideas.

After thinking about what the economy is, a natural next step is to ask: what is the best economic system we can design?

I landed on that question through another question: can advances in computation make socialism feasible? I went down this rabbit hole for my bachelor’s thesis.

The answer I found was: computation could solve some of socialism’s problems, but it couldn’t make it feasible, as innovation requires competition and profit incentives, which are incompatible with public ownership of the means of production.

During my master’s, I spent some time researching about economic theories of innovation, but the field was too hard to contribute something meaningful to.

Then one day I was having dinner with a friend, he started talking about how to solve the problem of economic inequality, and I thought: why not charge lower prices to poor people?

Imagine one bakery and two customers: a rich one who can pay 15 for bread, and a poor one who can pay 5. With a single price, the bakery earns more by selling only to the rich at 15 than by selling to both at 5.

There is a Pareto inefficiency: if the bakery could also sell to the poor at 5, it would earn 20 instead of 15, and no one would be worse off.

For my master’s thesis, I wrote a Python program that simulated what would happen to different market structures if there were more price personalization, and concluded that some types of markets could become less competitive.

After my master’s, I worked as a research assistant for Prof. Sanchez-Pages while I decided whether to do a PhD.

Yet I couldn’t stop thinking about my master’s thesis: with all the data there is about individuals, why aren’t prices more personalized?

I thought about turning this into a startup, which seemed like a strange idea: I didn’t know anyone who had founded a startup, and it wasn’t a common path after studying economics.

Nonetheless, I pitched the idea to my flatmate at the time, and he was hooked, so we started building Sumtive.

Our product allowed shoppers to anonymously share their bank transaction, social media, or Google data with the online shop they were browsing, in exchange for a personalised discount. An AI model estimated how likely each shopper was to buy, and set their discount accordingly.

The vision was that sharing your data once, in any partner store, gave you access to discounts in all of them. We turned this idea into a standalone app, where shoppers could find every partner’s discounts in one place: an AI-first shopping assistant, combining a chatbot and a personalised infinite-scroll feed.

We fundraised in San Francisco and launched on the UC Berkeley campus, and I fell in love with the act of building.

In the end, in-store personalisation was a nice-to-have for shops, not a burning problem: it boosted conversion slightly, but stores cared more about bringing traffic than converting it. Our app was meant to address that, but competing with Instagram and Pinterest required either raising a lot of capital or going viral, and after some painful months, we couldn’t achieve either, and ran out of money.

I landed back home at my parents’ house as my co-founder moved on to become a literary agent.

For a while I went co-founder dating looking for people who had an insight on a concrete, validated pain point that could be solved with AI.

I started working with someone on an AI automation platform for a specific accounting workflow he manually did at his job.

Since he was the one who knew about accounting, he would specify how the product should work in a document, and I would write the code and update the document.

I thought to myself: wouldn’t it be great if the document was linked to the code? He could use AI to edit the document, and the code would change to match. And whenever the code changed, the document would update too, so it always explained exactly what the product did.

Let’s call the document a Nota! As the accounting idea faded away, I decided to build Nota on my own.

The product allowed non-technical people to collaborate in the process of building software by editing the Nota through a web interface.

The idea turned out to have little defensibility: design partners wanted to run it on their own Claude Code or Codex subscriptions rather than pay my API prices, which meant open-sourcing an idea that was easy to copy in the first place.

I’m still glad I worked on this project. Not having a technical co-founder made me go through some painful mistakes, yet working through ideas yourself is the only way to learn how to think.