← All projects

iOS · Android 2026

BerichtsheftKI

AI-written training reports

Every apprentice in Germany has to keep a weekly training report, and almost every apprentice hates it. BerichtsheftKI fixes that: describe your week in a few words and the app writes a clean, submission-ready report.

Built end to end: app, backend, AI pipeline and the boring paperwork formats that make instructors happy.

Status
Live
Platform
iOS and Android
Role
Design, engineering, operations

Case study by Fabian Bitzer 5 min read

BerichtsheftKI is an app for apprentices in Germany that turns a short voice message about their week into a finished training report, the Berichtsheft. It knows the official training framework of more than 350 occupations, exports compliant PDF templates and runs on iOS, Android and the web. I built it myself, from the first idea to running it in production.

What a Berichtsheft is and why it needs an app

Apprentices in Germany’s dual vocational training system have to keep a training record, known as the Ausbildungsnachweis or simply the Berichtsheft. Week by week it documents what happened at the company and at vocational school, and the chambers of commerce (IHK) and of skilled crafts (HWK) require it. Without it, apprentices are not admitted to their final exam, as set out in section 43 of the German Vocational Training Act.

In practice, the Berichtsheft is the task almost every apprentice puts off. After a long working week there is no energy left, the right wording does not come, and weeks or even months pile up that have to be written in a rush just before the exam. That is where BerichtsheftKI comes in: writing the report should take no more effort than sending a voice message to a friend.

How writing a training report with AI works

The flow is deliberately short. Four steps turn into a finished weekly report that can be printed and signed.

  1. Set the occupation once: The app knows more than 350 apprenticeship occupations across industry, trade and skilled crafts. The AI writes in line with the official training framework of that occupation and in its professional vocabulary.
  2. Talk about the week: 45 seconds are enough to describe what happened. Anyone with more to say can record for up to two minutes.
  3. Let the report be written: The recording becomes a well-phrased, structured report in the tone trainers expect.
  4. Review and export: Adjust the text, export it as a PDF, print it or pass it on. Done.
The app at a glance: the dashboard with the week, the recording with its guiding questions in the middle, and the finished weekly report with hours per day. A 45-second recording is enough.

The difference from a general-purpose AI chat: BerichtsheftKI knows which content each occupation requires, knows the usual formats and does not need the context explained again for every report.

Three compliant report templates

Companies and chambers accept different formats, so the app exports three templates based on the BIBB model forms 2a, 2b and 3b published by Germany’s Federal Institute for Vocational Education and Training:

Template Format BIBB model
Weekly report Continuous text per week 2b
Daily report Table from Monday to Friday 2a
Weekly report with framework Activities mapped to the learning goals of the training framework 3b

The PDFs come in a print font or a handwriting-style font, optionally with the company name in the header. Which template is mandatory is decided by the company or the responsible chamber, so a quick check with the trainer is worth it.

The export page of the web app: one of the three BIBB templates is chosen, then the font, and the preview sheet follows both. The weeks shown are made up.

More than writing: features for everyday training

A good digital training record does not just help with wording. It also helps apprentices stay on track in the first place.

  • Express mode: Anyone weeks or even years behind can generate every missing weekly report for a period and a department in one pass.
  • Importing old reports: Reports written earlier can be scanned in as a photo or PDF and digitised.
  • Streaks, reminders and an exam countdown: Configurable push reminders and a countdown to the exam keep the routine going.
  • Quick actions: Sick, on holiday or a public holiday? One tap and the week is done.
Express mode in the web app: a period is set and every missing weekly report is generated in one pass. The weeks shown are made up.

Privacy: what happens to the voice recording

For an app used by young people at work, privacy is not a side issue. The voice recording is deleted right after processing, and only the finished text is stored. Every AI feature requires a signed-in account, and in the database each user’s data is separated with row level security. The details are in the BerichtsheftKI privacy policy (German).

The path of a voice message. Every AI call runs through an edge function that checks the sign-in, the recording is deleted right after processing, and only the finished text is stored, separated per user with row level security.

Platforms and pricing model

BerichtsheftKI is available for iPhone and Android, and as a web app in the browser. The first iOS version has been in the App Store since April 2026.

The first six reports are free. After that there is a monthly or yearly subscription, each with a free trial week, or Express mode as a one-time purchase. Current prices are shown in the app.

The web app dashboard: a saved draft is reviewed and finished, and the week and the count of finished reports follow.

How I built the app

BerichtsheftKI is a solo project: concept, design, development, store releases and day-to-day operations are all mine. The publisher in the app stores is BLUNATECH GmbH.

The app

The app is built with Flutter, one codebase for iOS and Android. App state is managed with Riverpod, navigation with go_router. The PDF templates are generated on the device itself, which keeps exports fast and independent of the server.

Backend and AI

Supabase runs in the background: a Postgres database with row level security, sign-in via an email link, and edge functions for every call that must not happen on the device. Speech is first converted to text by a dedicated speech-recognition service. A fast language model then writes the report, connected through OpenRouter, which makes switching models and tracking the cost of each request straightforward.

Payments, web and operations

Purchases and subscriptions run through Apple and Google and are managed with RevenueCat. The web app is built with Next.js, the website with Astro. I monitor errors and usage with Sentry and PostHog, and I automate releases for both stores with fastlane.

What I learned building it

Voice beats forms. An empty text field is the biggest hurdle with a training report. A 45-second recording, by contrast, is done in no time, even on the train home. That is why the whole app is built around voice input.

A 200 status code does not mean the answer is right. Models with a reasoning phase can spend their entire token budget on thinking and then return nothing useful, even though the request succeeded. The lesson: the budget has to cover both thinking and answering, and the choice of model needs regular review.

Access rights have more than one dimension. A trial week, a subscription and a one-time purchase can all exist at the same time, and a single record per user cannot represent that cleanly. Today access is evaluated separately per feature, and every AI endpoint checks the sign-in.

Try BerichtsheftKI

Information, guides and support are on berichtsheftki.de (German). Do you train or mentor apprentices and have feedback on the app? Write to me at kontakt@bitzer-fabian.de.

Next project

MMK Chatbot

Study questions for DHBW Mannheim, answered from the official documents.