Case studies · English

What we installed, and for whom.

Four cases in three beats — the situation, what we installed, what changed. No figure we cannot show. Claude is the platform we build with every day: Claude Code, the Claude API and MCP, Skills.

Higher education · since 2022

OMNES Education — AI on the admissions funnel of a 13-school group

The situation: A French higher-education group (13 schools, ~175,000 applications a year) with an acquisition funnel — prospect, applicant, admitted — spread across marketing, admissions and IT, and multiple tools and teams. DL leads the product side of this funnel since 2022: three critical applications, a team of seven (product owner, developers, QA, support), and a governance shared between the growth, marketing and IT directions.
What changed: A relationship that has lasted four years and widened every year, from product management to AI initiatives. We publish time or conversion figures only once they have been measured; the group's numbers are its own.
What we installed: Automated checking of supporting documents by AI
AI agents on WhatsApp for prospect conversations, written back into the CRM
AI-generated candidate profiles sent to partner companies for apprenticeships
The applications themselves are developed with Claude Code; the team's tools (Atlassian Jira and Confluence, among others) are connected through MCP
Modernised front-ends, application forms, API platform and observability

Training organisation · Île-de-France · 2026

Drainage Academy — a training school gets its admin back

The situation: A vocational school for lymphatic-drainage practitioners, one founder, in-person sessions funded by the French CPF scheme. Every learner meant a set of regulatory documents produced by hand — training agreement, attendance sheet, attestation, completion certificate, self-assessment — copy-pasted from Word templates, and prospects scattered across email, SMS and WhatsApp.
What changed: Document preparation went from manual copy-paste per learner to a few clicks, with a sending history the founder can trust. Hours saved will be published once measured with him — not before.
What we installed: A private management space (Next.js · Supabase · Vercel) built with Claude Code: a CRM with each contact's history, session management, and one-click generation of the five documents, faithful to the school's own templates, sent as branded emails with a visible sending history
A review-request email at the right moment, and a knowledge base that will feed the school's assistant
On the existing Laravel website: SEO foundations, a config-driven system of local pages (“training in <city>”), a city-filtered agenda, the official CPF badge, video testimonials, a persistent WhatsApp button
DL skills written for the job and reusable for any training organisation: reproducing Word templates as generated documents, safe deployment to a shared host, industrialised local pages

Premium packaging integrator · Tunisia · 2026

KARTHA — a brand, a site and a content engine before the first customer

The situation: KARTHA turns anonymous bulk olive oil into a finished product branded under the producer's own name: custom Bag-in-Box and cartons, QR traceability, one contact from sketch to pallet. An early-stage venture DL co-develops, pre-revenue — we say so.
What changed: When a producer asks ChatGPT who does premium packaging for olive oil in Tunisia, KARTHA is among the answers — the “be found on ChatGPT” offer, installed before we sold it.
What we installed: A complete brand dossier: positioning, customer research, competitive analysis, a 12-month marketing plan, content strategy, art-direction brief, design system
The website (kartha.fr), built on Lovable and deployed through GitHub with Claude Code — no platform credits — with an SEO/AEO foundation so that AI answer engines can read it; an internal cockpit for contacts and quotes
A social-media pipeline (self-hosted Postiz + front-end): drafts generated with Claude models from a versioned brand skill; every post validated by a human before publishing

Our own studio · 2026

Running a two-person AI studio on Claude

The situation: Our own operations — strategy, brand, sites, prospecting, watch, client follow-up, admin — with a small team. Everything we sell, we run on ourselves first.
What changed: Governance we apply to ourselves: AI proposes, a human validates every send and publication; personal data anonymised in every deliverable that leaves the studio; every deliverable registered and findable. In four weeks: 252 registered deliverables, 207 training organisations sourced and scored, a fixed positioning, a redeployed site.
What we installed: Claude Code as the workshop, connected through MCP to the tools we actually use: Airtable CRM, Gmail, Google Drive, Lovable, Firecrawl, Qonto, Notion, Todoist, Playwright, a live browser session
30+ versioned skills in a private repository — design system, client follow-up template, document generation for training organisations, site deployment, local SEO pages, brand skill packs — each with a learning loop so the skill improves after every use
An always-on agent (Hermes Agent, open source, on a VPS) bridged to Claude Code through a private Git repository — task inbox, results, shared board: overnight watch, prospect verification (207 training organisations verified against the CRM's exact vocabulary in one night), weekly email reports in a fixed readable format
A cockpit page generated from data: decisions to take, actions by owner, datasets — one view of what the team produced and what is left to decide

Talk to us How we work with Claude →Version française