It was on the roadmap:
"We need to aggregate Carif-Oref records on Réfugiés.info before March."
Fine.
But a question quickly emerged:
What if we built a real AI-assisted writing tool instead?
Not a demo. Not a hacked-together proof-of-concept.
A useful, robust tool designed for Réfugiés.info's editorial team.
The need is simple: we receive raw data (often from Carif-Oref), and before publication, it must be understood, simplified, rewritten in clear language.
This meticulous, human work takes time — lots of time.
Hence the idea of a writing copilot:
a tool that ingests data, proposes an initial rewrite, and lets humans correct, adjust, and validate.
An AI + human workflow,
where humans always stay in the loop.
The Hackathon — Two Days to Start Over
We launched a two-day internal hackathon with everyone in the same room: product, intrapreneurs, editorial.
We prompted, tested, explored.
At that point, we were still working with N8N. A good orchestrator, but quickly limited when you start dreaming of a tool that actually reasons.
We built prompt chains, classification flows, attempts at automatic rewriting.
Lots of ideas. Lots of energy.
But not yet the path.
After the Hackathon — Discovering Letta
I spent a few days exploring other options.
And I stumbled upon Letta.
Letta is a system for creating "stateful" AI agents — capable of memory and persistent context.
The opposite of "stateless" agents that forgot everything with each request.
What does that change?
Everything.
For the first time, we could imagine an AI capable of following a real editorial process:
import → filter → rewrite → export.
Without starting from scratch at each step.
That's that rare moment when you feel the missing piece click into place.
We finally had a solid foundation.
Nour, the Prompt Master, and the Frustration of "Almost"
Meanwhile, Nour, our intrapreneur prompt master, kept exploring.
With patience and talent, he was getting stunning results.
But there was always this "almost":
the model rewrites too much, sometimes hallucinates, loses an essential nuance.
And that "almost" is exactly where everything tips toward doubt for the editorial team.
With Letta, we hope to reduce these drifts:
an agent that keeps its state, follows a framework, stays coherent from step to step.
An AI that's less brilliant but more reliable.
And for the editorial team, that's exactly what we need.
And Now?
The exploration phase is behind us.
We're moving to the building phase:
👉 a POC in two weeks,
👉 a complete version by March 2026.
From today on, I'm sharing behind-the-scenes of the project here:
the progress, the choices, the failures, the wins, and the small human moments.
Because it's often in the details and the fumbling that true adventures are born.
There we go.
This is officially day 0 of building in public our augmented writing tool.
Jérémie Gisserot — Developer at Réfugiés.info