Starting Fresh, Embracing Clarity
Today, we made the decision that conditions everything else:
👉 we're not rebuilding on top of our old editorial back-office.
We're starting from scratch.
Meeting this morning with Luis (CTO) and Julie (PM). We asked the hard question: "Do we patch things… or do we commit to a real tool designed for AI, editorial work, and longevity?"
Unanimous answer: we go for it.
Our current back-office, despite years of service, has become a mess: implicit dependencies, tangled features, technical debt. Impossible to build a clean AI workflow on top of that.
And most importantly: we want to isolate the AI flow (import → filter → rewrite → export) to never mix generated content with final content.
What We're Building (and For Whom)
The goal: an editorial copilot for Réfugiés.info.
It takes raw data (like Carif-Oref for training programs), classifies it, rewrites it in clear language, then lets humans decide.
Our mantra: AI assists, humans decide.
We're speaking to three audiences:
- Intrapreneurs → vision, impact, public service integration
- Product owners → value, scope, milestones
- Developers → architecture, stack, concrete constraints
The Roadmap (Realistic, but Tight)
- Import (Carif-Oref first)
- Filter (sort, classify, detect duplicates)
- Writing assistance (clear language + human validation)
- Export (to main database, draft side)
Goal: a POC in 2 two-week sprints.
Then a solid version by March.
Our key metrics: filter quality, rewrite reliability, correction rate requested by editorial, actual time saved.
Tech: Our Deliberate Choices
- Letta (stateful agents) → persistent memory, fewer hallucinations, ability to work on a record over time.
- ADE Letta → edit and version specs continuously.
- Next.js + Tailwind + DSFR → "public service" consistency, shipping speed, accessibility.
- Drizzle ORM → typed, readable, efficient in monorepo.
- BlockNote → a rich editor suited to the "AI proposes / human adjusts" workflow.
- SpeckKit + Windsurf → spec-first mode: we design before coding, AI executes.
Since we've been working this way, we're less "dev vibes", more architects.
Field Feedback: AI Must Stop Making Things Up
The editorial team knows exactly what it wants — and what it refuses:
- a tool that talks down to them,
- rewrites that sound good but are false,
- hallucinations dressed up as certainties.
Nour is already working wonders with his ninja-level prompts, but we need the AI to be more reliable and learn from continuous feedback.
Letta can change that: keep a state, a context, a file.
We move from an AI "that answers" to an AI "that works".
Today's Alignment: What We're Locking In
- ✅ Tool separate from production, fresh database
- ✅ Master flow: import → filter → rewrite → export
- ✅ Validated stack: Letta, Next/Tailwind/DSFR, Drizzle, BlockNote, monorepo
- ✅ Method: spec-coding (SpeckKit + Windsurf)
- ✅ Milestones: POC in 2 weeks, solid version by March
The Small Human Moment
There was this silence, right after the question "do we keep the old system or not?"
The comfortable choice would have been to patch what exists.
The right choice was to clarify, start on solid ground, and really align.
It's simple to write.
In reality: it takes shared courage.
And Now?
Starting tomorrow:
- validate the technical foundation,
- wire up Letta,
- import a first Carif-Oref sample,
- test the filter (no mercy),
- integrate BlockNote to validate the "AI proposes / human adjusts" flow.
Building in public starts for real.
See you tomorrow.
Jérémie Gisserot — Product Designer & Developer, Réfugiés.info