Content integrity in the age of generative AI
A custom detection system that tells student-written notes from AI-generated uploads, protecting a global knowledge-sharing library.
Result
Twice the flagging rate of the industry-standard detector, detecting up to 98% of AI- and user-generated content in documents.
The problem behind the brief
Studeersnel's business depends on students sharing authentic study material. Large language models created a new risk: a surge of AI-generated uploads that would dilute the quality and reliability of the library.
To keep users' trust, Studeersnel needed to identify where every document came from. Off-the-shelf detectors weren't precise enough for academic writing, and the system had to keep up with a very large daily upload volume.
How we approached the build
One team covered strategy, data, ai engineering, working from the same decisions from the first assessment to launch.
Strategy
The work focused on technical de-risking and data quality. A detector only works if it understands how student writing differs from machine output, so the plan centred on generating high-quality training data before the problem became mainstream.
Data
The team built a data structure to generate AI content in bulk and assembled a training set that allowed precise comparison between human and machine text.
Engineering
The team fine-tuned the RoBERTa language model for detection in an academic context and integrated it into Studocu's existing pipeline, so documents are assessed in real time.
What it delivered
Performance
Twice the industry-standard flagging rate
The custom model detects up to 98% of AI- and user-generated content in documents, and around 80% in question answers.
Trust
Verified content authenticity
Knowing where each document comes from lets Studocu keep a library of authentic, student-written material, which is the core of the platform.
Capability
A head start on generative AI
Built before the public surge of generative AI, the system let Studocu scale without compromising content integrity.
“Miyagami helped us develop a high-performing AI model. Seeing the final product work so effectively has given us the perfect head start.”
Get a shortlist within five working days
You share the roles and the stack in a short form or a thirty-minute call. Within five working days you get named senior engineers to review, each with both scorecards.
ISO 27001Certified