Career Journey
Path, tools, things I've built
Path
At Bocconi I discovered methods for validating an idea before building it. It seemed obvious, but it isn't. I applied that method to my first project — an AI tutor — and I realized that doing field research is much more useful than sitting around thinking. I really enjoyed design thinking and the scientific method.
Two internships: one at an incubator, one at a university research center. I talked with dozens of founders and investors. The most useful thing I learned wasn't how to build a startup — it was understanding how not to, and which cognitive biases lead almost everyone to make the same mistakes.
Joined Nana Bianca. For the first few months I did what every newcomer does: listened, observed, tried not to say anything dumb. It worked well enough. I met a lot of smart people. After a few months I took an SQL course because I thought I wanted to become a data analyst.
I started giving startups my opinions and talking with VC funds. Meanwhile, I took a product management course because I thought it could be a nice way to get closer to the product-user interaction. I was starting to get interested in user research, analytics, and testing.
I started using AI the day ChatGPT came out. For a long time I tried pretty much every tool that got released along the way. I have to admit that for about a year now, the way I use it is much more actionable. In fact, since 2025 I've thrown myself into building things.
Here are my first experiments thanks to AI. A local personal finance dashboard to manage my money. A mentor platform for Nana Bianca built with Bubble (a no-code dev tool). A copy-automation workflow from video for a podcast with n8n. In the meantime I also co-organized the first official Cursor meetup in Italy — why not.
Thanks to my experiments and my drive to build a product, I was handed the development of Dora — an AI assistant for online shopping. There was just one small problem: the dev who was supposed to build it had already left before starting. So for the first month I built a prototype myself with n8n and Cursor, tested with real users. In December, three developers joined.
Full-time PM on Dora. After work I'm a co-founder of Uilo, a guidance tool for high school students, launched in March 2026.
Experiments in depth
Dora prototype — AI shopping assistant
October 2025
They moved me into a new role to work on an AI assistant for online shopping. A brand-new project, no defined process, and after a week the dev who was supposed to build it quit before even starting.
So I built the prototype myself.
I made some hypotheses about the intentions of people who shop online — what they look for, how they phrase their questions, what they expect as an answer. From those hypotheses I built the prompt, the flow, and the backend with n8n using the Agent feature. With Cursor I built the frontend in vanilla HTML, CSS, and JavaScript. In a few weeks I had something working to test.
I collected informal feedback and ran 10 live tests with real users. Useful for understanding what worked and what didn't. Then I hit n8n's limits for the later iterations and handed everything over to a dev to build the MVP.
Now Dora has a team of three developers and I run it as a full-time PM.
Nana Bianca mentor platform
January 2025
Every time an acceleration program ended, the mentor payment process was a mess. Endless email threads, hours lost tracking down scattered information. Nobody had the time to fix it.
So I did it myself, after work, in Bubble.
Four types of login (admin, mentor, startup, partner), mentorship tracking, billed hours, amounts, who to bill and how. All in one place. The finance team now checks the hours on their own — zero email back-and-forth between the acceleration team and the finance team.
It's ugly, I know. But it was January 2025 and I didn't yet know how to use anything else. Convoluted enough to maintain internally, but a year on I have to say it got used and did a decent job. We later retired and replaced it.
Figuring out whether to implement in-platform booking of mentorship hours. At first I dove into managing the automation flows: send request by email, pending status, acceptance or proposal of a new date, handling the counter-proposal, etc. Only later, talking to some mentors, did I realize it made no sense. Almost all of them already use their own calendar with availability. Since we weren't interested in storing the calendar in the platform, it was enough to let the mentor add a link. Lesson learned: talk to users before building the flows.
Prima o Poi podcast: automation attempts
January 2025
I curated the weekly rundown of the Prima o Poi podcast on AI, innovation, and business. I read sources, picked topics. A manual, repetitive task — an obvious candidate for automation.
I built an n8n flow to automatically research topics. It didn't work — it gave me exactly the same articles I would have found on my own in five minutes. I even added scoring logic to filter the results. Dropped that too. It took too much time to produce something valuable compared to doing it by hand.
What did work was something else: a flow to extract the best snippets from the full video to use as shorts. The flow worked like this: the transcript came in, n8n split it into chunks, extracted a list of candidate snippets with a title and timestamp, generated 5 pieces of copy tailored per platform, and uploaded everything to Notion automatically. Something that initially saved a lot of time. And honestly, today a single prompt to Claude is enough.
In my spare time I was also asked to build an analytics dashboard to monitor the podcast on YouTube, Spotify, and Apple Podcasts — and then ideally the social channels. I tried building it with Cursor using the available APIs. YouTube worked well. With Spotify and Apple I ran into API problems — they were either unavailable or only gave fairly shallow data. Almost never used — in the meantime they'd taken out a subscription to a tool that already did all of this and, on top of that, let you publish from one place.
It's wild to think that today all it takes is presenting a problem to Claude, giving it access to the right tools, and in a few hours your solution is live.