Syrtis — an AI agents platform
Co-founder of an AI company: a generic platform for configuring agents as graphs, its whole infrastructure, and the client projects built on it — more than ten thousand commits in two years.
Syrtis began as a learning platform, meant to adapt the transfer of knowledge to each learner. It grew into something more general: a system for configuring AI agents. A scenario is a graph of nodes, some thirty types of them: model calls (reasoning, structured output, tool use through MCP servers, image generation, speech, vision), sandboxed Python, conditions, iterators, sub-jobs, SQL, Cypher and HTTP requests, n8n workflows, documents for retrieval. Nodes exchange tagged messages; every scenario is versioned on save and travels as JSON; a visual editor lets you follow a session message by message, down to the prompt each model received.
Underneath: a Python engine driven by RabbitMQ jobs, a Symfony API, a Symfony and TypeScript editor kept live with Mercure, PostgreSQL, MongoDB and Neo4j for the knowledge graph, OpenAI, Anthropic, and Ollama for local models, client SDKs in JavaScript, PHP, Python and Flutter. Around it, everything the company runs on, self-hosted: a Matrix chat, a Nextcloud drive, GitLab and its Docker registry, monitoring, n8n, Odoo.
Then the projects of our clients, built on the platform: for FIBA, an API writing the report of a basketball game from its statistics and play-by-play — scoring runs, lineups, player efficiency; for an audit firm in Luxembourg, an assistant drafting engagement letters inside their application.