---
title: "Before all this: a multimodal assistant for superyachts, and Pepe AI, a tutor that closed."
url: "https://simonebova.com/en/runs/prima"
lang: "en"
author: Simone Bova
alternates:
  it: "https://simonebova.com/md/it/runs/prima"
  en: "https://simonebova.com/md/en/runs/prima"
  pl: "https://simonebova.com/md/pl/runs/prima"
---

# Before all this: a multimodal assistant for superyachts, and Pepe AI, a tutor that closed.

A voice and vision assistant for a superyacht builder, training junior developers, and Pepe AI, a university tutor that closed in 2026. What each one taught me.

Name: Before all this

Terrain: Software house · EdTech

When: 2023 → 2026

Role: Developer, then CTO

| Fact | Value |
| --- | --- |
| Terrain | Software house · EdTech |
| Superyacht | Team of three |
| Pepe AI | Incubated at I3P |
| Outcome | Closed in 2026 |

## The situation

For two and a half years I worked part-time at a software house. The biggest project was a mobile assistant for a superyacht builder, with voice, text and vision in the same app. At the end of 2025 I joined Pepe AI as CTO: the university tutor from SpeedingCorsi, incubated at the Politecnico di Torino's I3P. The MVP was already live, and my job was to turn it into a product that holds up.

## What I did

On the superyacht assistant I handled requirements, backend and part of the project management, delivered the piece that recognises the voice and who is speaking, and led the three-person team on the mobile version. Alongside it I worked on migrating an old CRM, from the interface concept to the accessibility guidelines, and trained small groups of junior developers. At Pepe I took over product and technology.

## What changed

What I carry from those years is mostly a way of training people and a couple of judgement errors I wouldn't repeat. Pepe closed in 2026. We had a product that worked, and traction and budget weren't enough to keep going. There's no single technical cause, and nobody to blame.

## What I learned

In the superyacht app I delivered the piece that recognises the voice and who is speaking. “Multimodal” gets sold as a model capability. The real work was keeping three questions apart: what this voice is saying, who is speaking, what this person is allowed to do. Answering the first one well tells you nothing about the other two, and a demo usually shows only the first. Years later, on CVing, I applied the same separation to agents on WhatsApp: the channel says who is writing, and a separate check decides what they may do.

At the software house I trained small groups of junior developers. Each group got a bounded outcome, with a real user, and carried it from requirements to deploy. Training plans are usually split by language or framework, so a junior learns to write code without ever seeing the decisions that come before and after it, the ones that make software usable: what the user is asking for, what data is needed, who can actually use the interface. Today, when I review someone's work, I look at those decisions before the code.

On a project from those years we had to show certain documents only to certain people, based on their role. Everyone around me treated it as a huge problem, and I believed it too. Months later I saw much larger systems handle permissions down to a single field, and realised the difficulty lived in the environment I worked in. I used to think the people around you shape your priorities. They had also shaped my sense of what was possible. Now, when a problem feels impossible, I go and look at someone who solves a harder version of it.

> I fell for the trap.

SpeedingCorsi started as live tutoring for university students. When I joined as CTO, Pepe's AI tutor was already live: it walked students through exercises step by step and asked questions to check they had really understood. Students, subjects and the problem were already there, so moving from service to software looked low-risk. Buying a course with a person and paying to come back to an AI tutor every week turned out to be two different exchanges. Since then I keep separate tracks for whether people like a product, use it every week and pay for it, and for each one I decide up front what proof has to arrive.

[Runs](https://simonebova.com/md/en/runs)

[simone@yempik.ai](mailto:simone@yempik.ai) · [Book a call](https://calendar.app.google/LKDrQPKGV2EAFCMN9)