---
title: "Assessing skills with AI: from Yempik's Assessment line to a micro exit."
url: "https://simonebova.com/en/runs/micro-exit-formazione"
lang: "en"
author: Simone Bova
alternates:
  it: "https://simonebova.com/md/it/runs/micro-exit-formazione"
  en: "https://simonebova.com/md/en/runs/micro-exit-formazione"
  pl: "https://simonebova.com/md/pl/runs/micro-exit-formazione"
---

# Assessing skills with AI: from Yempik's Assessment line to a micro exit.

With Raffaele Zarrelli we built Yempik Assessment and sold it as a micro exit to a training group. What the AI decides, and what it doesn't.

Name: Micro exit · training

Terrain: Startup · edtech

When: 2024 → 2026

Role: Co-founder: technology and production

| Fact | Value |
| --- | --- |
| Terrain | Training · HR |
| Team | Simone Bova and Raffaele Zarrelli |
| Origin | Yempik Assessment |
| Outcome | Micro exit, half mine |

## The situation

A lot of mandatory training exists to get a register signed. A certificate says a person was there, and what they understood stays unknown. Since 2024 Raffaele and I had been running a Yempik line on this problem, Yempik Assessment: AI to assess skills and guide upskilling, tested with companies in training and workplace safety and shown at ITCILO events. The platform starts from the trainer's materials, generates questions and scenarios, written or spoken, and returns results by skill area, so the trainer can see what to go over again in class.

## What I did

The product we sold comes from that shared work and two years of lessons. When a training group wanted it, we split the work cleanly. Raffaele designed the mock interfaces with the client and led the relationship. I took the technical side when it was time to land it in production: the verification before the handover, the infrastructure, the migration onto the buyer's accounts and every change all the way to production.

In the verification I started from a real, long course, and I held to one product principle: AI prepares the questions and explains the scores, while the outcome is computed by rules the trainer knows.

## What changed

In 2026 the platform was sold as an asset to a training group. It is a micro exit, and half of it is mine. Changes the client asks for reach testing quickly and production with the checks they need, and the move onto their accounts is in progress.

## What I learned

Before the handover I tested the platform with a real, long course, with an important topic on the last slides. A demo with a short course never tells you whether the assessment covers all the material: an accepted file and well-written questions make coverage look complete when it still needs checking. The same goes for a trainer updating a file after the questions were generated. Now, on any product that generates content from documents, I first check which sections that content comes from, starting at the end, and what happens when the document changes.

On the platform the model assigns a score for each objective and explains it. The system applies thresholds, weights and critical objectives, and generated questions wait until the trainer approves them, because a correct, clear question can still test the wrong thing. An assistant that converses with the worker easily ends up deciding the outcome as well. With the two roles kept apart, an outcome that has consequences for a person can be rebuilt step by step, without asking the model why it decided. Now it's the first boundary I set whenever a model assesses people.

The client chose to call the two assessments “initial” and “final” for workers, and “diagnostic” and “summative” for trainers and reports. Same concept, two names depending on who's reading. Consistent terminology usually means one word everywhere, and here it would have pushed the trainer's vocabulary onto a worker who uses a phone and sometimes speaks basic Italian. We kept both vocabularies in one place, so no screen, email or report invents a third. For that same worker, outcome colours always come with a word next to them.

*// name and figures stay with the buyer.*

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

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