Putting AI to work for the public interest
Discover how Marion Berthaut and the French Institute for Restorative Justice transformed 200 sensitive testimonies into actionable data for research and public policy, while preserving each person's dignity.
"We should start a project, but it would take me three months full-time. I don’t have three full months to dedicate to it."
NLP R&D Expert - Fractional R&D Lead For the past ten years, Marion Berthaut has been working in natural language processing (NLP). In 2015, she launched her first company to help people with dyslexia read more effectively, already using linguistic analysis technologies. Today, Marion works as a "fractional R&D lead." Her mission: spotting opportunities where others do not see them.
“You have an extraordinary body of textual material, of incredible richness. It just so happens that there are tools capable of leveraging this textual matter especially when there’s a lot of it and extracting some really fascinating insights from it.”
Marion needed an “off-the-shelf” solution, quick to deploy, without having to build everything from scratch.
Data hosted in France • Full control over inferences • No training on third-party models
Rapid experimentation • Frictionless trial-and-error • No benchmarking for each new idea
Ready to use solution • No development from scratch • Production deployment in a few weeks
Access to Mistral, Anthropic, and others • Cost optimization • Selecting the best model for each use case
"For ideological reasons and the values at stake, I wouldn’t use certain models in the context of restorative justice. The idea that testimonies could be stored on servers outside of France was simply inconceivable."
Marion discovered Swiftask at the beginning of 2025 thanks to Olivier La Grandeur. At that time, the AI solutions market was still young. “You have to put yourself in the context—it’s January 2025. There aren’t 36,000 solutions available,” Marion recalls.
Data Exploration for Restorative Justice – Digital Intelligence
A tribute to Edwin Lemert, who theorized the distinction between primary and secondary deviance.
Phone interviews are transcribed automatically with high accuracy, preserving nuances and emotions.
An agent detects and removes all personal identifiers (names, locations, dates) based on the standards of the Cour de Cassation.
Once anonymized, the testimonies can be provided to research laboratories for in-depth analysis.
A Pragmatic Approach
Marion released the first version of Edwin in May 2025, working with 50 testimonies. She tested different models (Mistral, Anthropic) and optimized costs by using models that consumed fewer tokens.
An ROI measured not in hours saved, but in value created
"It’s a game changer. It allows us to do things that weren’t accessible before. Edwin saved us time, but more importantly, it made us decide to act. Where the time required was prohibitive, we simply wouldn’t have done it."
Major Discovery
Edwin detected a particular anger from victims toward the penal system itself, revealing a systemic issue that had not previously been documented through the testimonies.
Giving Participants a Voice
IFJR was able to provide the Ministry of Justice with insights that no one else could offer during public policy evaluation: the voices of offenders and victims.
Funding Secured
€150,000 in funding was secured to develop Edwin 2. The POC served as a proof of concept for private foundations. Before Edwin, IFJR would have simply “not applied for the grant.”
Game Changer
Edwin enabled the decision to act. Where the time required had been prohibitive and so nothing was done it became possible. This represents a scale-up opportunity for IFJR.
.png&w=48&q=75){kind=small}
.png&w=48&q=75){kind=small}
Democratize access to testimonials for research
The next step
Now that Edwin is trained, Marion is waiting for the €150,000 funding to be released to further develop Edwin.
"An agent will return the intelligence you give it. So if you ask it stupid questions, it will answer stupidly. That’s how it works: it amplifies. If you amplify zero, well, it will still be zero. But if you amplify intelligence, then it’s astonishing."
Open to a small group of IFJR members for daily use
Supported by training to learn how to ask the right questions
Accessible to partner research labs with strong security guarantees
"There are tools around language processing. These tools are super powerful. It's incredible. Rather than just using them to analyze online reviews to figure out which second toaster you’re going to buy when you don’t even need one..."
"What if we put that power to work on something that benefits society?"
Testimonies collected since 2018
Vulnerable populations
Sensitive data protected
12 employees
46 boulevard d'Alsace Lorraine - 64000 PAU (Pyrenees-Atlantiques)
Discover how Swiftaskcan multiply your team's productivity
