ArnaudGuiovanna — AI Product Builder, LLM-based Intelligent Tutoring Systems, Normandy, France

AI Product BuilderLLM-based Intelligent Tutoring Systems

I build AI tutors and the pedagogical harnesses that make them reliable.

Based in: Normandy, France
Experience: 10+ years of digital transformation in the training sector

Projects

Active

tutor-mcp

An open-source MCP server that turns any LLM into an Intelligent Tutoring System. The AI creates and explains; a persistent, auditable engine decides what to learn, practise or revisit next.

Open source, MIT, alpha

Generative teaching, deterministic pedagogy: The AI builds the curriculum and teaches; the engine governs progression and keeps the evidence.
Learning engine: Bayesian Knowledge Tracing, FSRS spaced repetition, prerequisite rules and assessment evidence, with narrative memory across sessions.
Tutor ACP: A Windows client that connects Codex or Claude to an institution's learning service and its shared catalogue of programmes.

v0.7.3server0.0.4Tutor ACP3profiles: institution, local, hobby

Go, SQLite, PostgreSQL, OAuth, any MCP client (Claude, ChatGPT, Gemini, Le Chat)

Website GitHub README

In development

Multi-agent system for Moodle

A multi-agent system for a national training platform, integrated with Moodle. Three specialised agents (adaptive, classify and generate) collaborate to handle requests about learning pathways, teaching resources and their classification.

Professional project, closed source, 2025 to 2027

Funded by France 2030

Adaptive agent: personalised pathways: Resources selected from the positioning test, then targeted remediation.
Generate agent: trainer assistance: Teaching sequences assembled from existing resources.
Classify agent: content classification: Richer metadata so documents and media are easy to find.

Multi-agentorchestrationMoodleintegrationOct 2027go-live

Java, Spring Boot, LangChain4j, PostgreSQL, Redis, RabbitMQ, Docker, GitLab CI/CD

Building

Ouvrier

The general-purpose harness framework behind my approach. A Go framework for agentic middleware: declare a trigger, LLM-powered pipes, their tools and the outcome, and Ouvrier runs it as a governed worker.

Open source, MIT, v0.1

Runtime: HTTP, cron, webhook and stream triggers through one harness.
Governance: Permission policies, sandboxed tools, hooks and traces.
Outcomes: Typed results, schema validation and durable state.

4triggers3providersGo 1.25

Anthropic, OpenAI and Ollama adapters, SQLite state, sandboxed tool executor

GitHub README