Our stack

Fifteen interchangeable bricks. None imposed.

We work with these technologies. Each brick can be swapped without breaking the whole. If a better component ships in eighteen months, you integrate it in days — not in six months of rewrites.

Proprietary language models

Claude

Anthropic

Long reasoning, instruction-following, solid tool use.

GPT

OpenAI

Generalist, mature ecosystem, controllable cost.

Gemini

Google

Very long context, multimodal native.

Mistral

Mistral AI

European reference, EU hosting available.

Open self-hosted models

LLaMA

Meta

Wide range (8B, 70B, 405B), enterprise-usable license.

Qwen

Alibaba

Excellent multilingual, good size/performance trade-off.

Mixtral

Mistral AI

Mixture of experts, sovereign hosting possible.

DeepSeek

DeepSeek AI

Reasoning and code, very low inference cost.

Orchestration and agents

MCP

Model Context Protocol

Open standard for connecting LLMs and tools. Default recommendation.

LangGraph

LangChain

Stateful agent graphs, useful for complex workflows.

Temporal

Temporal

Durability and retry policy for critical pipelines.

Languages and runtime

Python

—

Data science, ML, integration scripts, pipelines.

TypeScript

—

Interfaces, APIs, client-side agents.

Docker

Docker Inc.

Containerization for reproducibility and deployment.

Kubernetes

CNCF

Orchestration at scale, often oversized for mid-caps — applied with judgment.

Databases

PostgreSQL + pgvector

—

Relational SQL plus vector search in the same database. Our default choice.

ClickHouse

ClickHouse Inc.

High-performance analytics on very large volumes.

Qdrant

Qdrant

Dedicated vector database when advanced features are needed (filters, payloads, multi-collection).

Sovereignty

Fully on-premise is doable.

Some cases require it: trade secret, sector compliance, defense classification, health data. We then deploy only open self-hosted models, on your servers or with a sovereign European host (Scaleway, OVHcloud, 3DS Outscale). No data leaves your perimeter. No public API is called. More demanding to set up, but it is the only option for some industries.

Discuss a sovereign deployment

A specific tech to validate?

We accept technical constraints from your IT or your CTO. Tell us what you already have, we'll tell you whether and how we can integrate cleanly.

Book a technical call