Home/Expertise/AI & Automation/AI assistants connected to your data
01 — Expertise · Generative AI

Integrate generative AI without dependency.

From scoping to deployment: RAG, model selection, security, costs, response quality, and reversibility.

AI assistants connected to your data
Your data
Internal documents
Database
Emails & tickets
Business procedures
Assistant
AI
Understands
Searches

Answers
Your teams
"What is the notice period in the Acme contract?"
30 calendar days from the date of notification.
Acme contract · Art. 8.2 · p. 12
Why us

Our way of working.

01

Independence

Multi-LLM by default. Llama and Mistral in open source, GPT, Claude, and Gemini in proprietary, depending on the task.

02

Pragmatism

We don’t launch a POC if the data isn’t ready. We don’t push to production if the evaluation doesn’t validate. No reckless forward momentum.

03

Continuity

AI must integrate with your tools, security rules, and operational practices. Not exist alongside them.

AI stack

The building blocks we integrate.

None of these are default choices. The stack is decided during scoping, based on your actual context.

Models

Llama
Mistral
GPT
Claude
Gemini

Frameworks

LangChain
LangGraph
LlamaIndex

Vector stores

pgvector / Supabase
Qdrant

Orchestration

Langflow
n8n
MCP
vLLM
Ollama
+ systematic benchmarking on your business corpus using Ragas, LLM-as-judge, and test sets built with your domain experts.
Client case

Sistr — AI Assistant.

Document search on anonymized patient files. Sovereign hosting on Scaleway, Llama 3.1 70B, RAG with source citations.

Read the full case
Healthcare · Public Sector2025 · 8 months

CHU Nantes

Semantic search across 12 years of anonymized reports, deployed for 1,500 practitioners and researchers.

−70%average documentary research time per practitioner
1 500active internal users · monthly
100%sovereign hosting, self-hosted open source model
Llama 3.1 70BpgvectorScaleway GPUvLLMHDSKYC
AI Method

A five-step project

A dedicated team from scoping to delivery with a dedicated project manager

01Scoping

Use cases & corpus.

Business feasibility, data quality, sovereignty constraints. Output: POC go/no-go decision.

02Model selection

Argued benchmark.

2-3 models tested on your real corpus. Open source vs proprietary decided on data, not preferences.

03POC

The core use case.

2 to 6 weeks. One use case, done right. We validate sources, rights, and assistant limitations.

04Evaluation

Quality metrics.

Accuracy rate, cited sources, hallucinations, cost per query, user feedback.

05Industrialization

Production deployment.

If criteria are met: deployment, monitoring, documentation, and handover to teams.

Sovereignty

Security, compliance, reversibility.

These topics aren’t addressed at the end. They frame decisions from the start: data, access, hosting, logs, monitoring, and exit clauses.

Sovereign hosting

Scaleway, OVH, self-hosted. Data and inference in the EU when required.

Open-source models

Llama, Mistral, Phi. Self-hosted via vLLM or Ollama. No data leakage to third-party clouds.

GDPR & compliance

Privacy by design. HDS-compatible for healthcare. DPIA support if needed.

Reversibility

Code delivered, documented, tested. No proprietary platforms. Everything is retrievable.

FAQ

Frequently asked questions

The most common questions during scoping. If yours isn’t here, reach out!

It depends on three criteria: data sensitivity, expected quality for your real-world task, and your tolerance for vendor lock-in. We systematically compare Llama/Mistral against GPT, Claude, and Gemini on your corpus before deciding. No model is universally superior—decisions are based on your metrics.

Yes. Three options: sovereign European hosting (OVH, Scaleway), self-hosted deployment on your GPUs, or local inference via vLLM. In all cases, the data flow is documented and auditable. For sensitive projects (healthcare, defense, finance), we work exclusively with self-hosted open-source models.

Three levers: (1) continuous evaluation on representative test sets, (2) systematic grounding (RAG or tool calls)—no free generation on facts, (3) safeguards (format validation, refusal in case of doubt, human escalation). No production deployment without measured and accepted quality metrics.

Three things are enough to start: a clear business use case in plain French, an executive sponsor, and access to relevant data. The rest—model selection, architecture, hosting—is part of our scoping process.

Yes, it’s one of our core services. Independent audit in 1-2 weeks, recovery plan, and code takeover. We precisely document what’s not working (architecture, quality, costs) before proposing next steps.

Always. Your teams co-build from the scoping phase and take over the code in the long run. We document, train, and organize a gradual handover. Our goal isn’t to become essential to your operations.

Yes, we offer technical training (LLM architecture, prompt engineering, evaluation) and business training (leading a use case, writing a product prompt). Formats: half-day, full-day, or multi-session programs. Catalog available upon request.
Get started

An AI assistant connected to your data to deploy?

Contact details
contact@agence-scroll.com
+33 6 48 03 90 27
20 Rue des Taillandiers
75011 Paris
Response within 24 business hours.