Case study/Education
Lifting meinUnterricht's search success by roughly 9%
unicrew embedded an engineer on meinUnterricht's search and discovery team, working the user-facing features on top and the AI and data layer underneath.
meinUnterricht is a German edtech platform that gives teachers access to a large, curated library of vetted teaching materials. unicrew embedded an engineer on their search and discovery team, working the user-facing side of search, relevance, filtering and autosuggest, on top of the AI tagging and embeddings underneath. A controlled experiment showed a roughly 9% lift in search success rate.
- Client
- meinUnterricht GmbH
- Focus
- Education, AI
- Market
Germany- Engagement
- Embedded engineer, search team
Outcome at a glance
What this project delivered, in numbers.
- ~9%Higher search success rateControlled experiment
The challenge
meinUnterricht set out to improve its search and discovery experience: better relevance, filtering and autosuggest, plus the AI tagging and embeddings behind a move to vector search. On a library of curated materials, search is the surface that decides whether a teacher finds what they came for, and a ranking that quietly buries the right worksheet does not show up as an error in a log. It surfaces when someone measures it.
Our approach
unicrew joined as an embedded engineer on the search team, over an extended engagement rather than at arm’s length from it. The work spanned the full stack of search: user-facing features on top, and the AI and data foundations underneath. unicrew scoped, built, shipped and measured features end to end, working directly with meinUnterricht’s product and engineering leads.
The AI layer was tagging and embeddings, the groundwork for meinUnterricht’s move to vector search. That is the kind of work our AI integration services and AI-readiness assessment are built to de-risk.
Results
The headline number here came out of a controlled experiment, which is why it is on this page at all. Three outcomes came from the engagement.
-
Search success rate: a controlled experiment showed a roughly 9% lift.
-
Shipped: the search features unicrew built are live in production.
-
The AI layer: the tagging and embeddings work set up meinUnterricht’s move to vector search.
Putting AI into a product that already has users is its own discipline: for another version of that problem, read AI-powered automation for project management, or hear it from clients directly on our client reviews page.
“They’ve delivered on time and responded well to our needs.”
Daniel Siebenson, Director of Product & Engineering, meinUnterricht GmbH
What the client says
Our headline result was a controlled experiment showing a roughly 9 percent lift in search success rate. The features built are live in production, and the AI tagging and embeddings work set up our move to vector search. What stands out most is their ability to own work end to end, from user-facing search features to the AI and data layer underneath.
04/Quick answers
The questions behind the project
What did unicrew do for meinUnterricht's search?
unicrew embedded an engineer on meinUnterricht's search and discovery team and worked the full stack of search: relevance, filtering and autosuggest on top, and the AI tagging and embeddings underneath. A controlled experiment showed a roughly 9% lift in search success rate, the features built are live in production, and the AI work set up meinUnterricht's move to vector search.
Was this an outside contractor or part of the team?
Part of the team. unicrew joined as an embedded engineer on meinUnterricht's search and discovery team, over an extended engagement rather than a fixed piece of work. unicrew owned features end to end, scoping, building, shipping and measuring impact, and worked directly with meinUnterricht's product and engineering leads.
What AI work sits underneath the search features?
AI tagging and embeddings. Alongside the user-facing work on relevance, filtering and autosuggest, unicrew built the tagging and embeddings layer that set up meinUnterricht's move to vector search. How that layer is built inside their product is meinUnterricht's to describe, not ours.
How do you tell whether a search change actually improved anything?
Decide what counts as a successful search before you touch ranking, instrument it, then run the change as a controlled experiment instead of shipping to everyone and reading the chart afterwards. That is where the number on this page comes from. If a product cannot hold part of its traffic aside, or if the definition of success moves with the release, the resulting figure will not survive the first person who asks how it was measured. Relevance is easy to feel and hard to prove, so the measurement design deserves as much attention as the ranking change.