Gianluca Mazza — AI Systems Engineering

Designing production AI systems with agents, tools, memory, retrieval, evals, durable runtime state, and recoverable execution paths.

Systems

Multi-agent runtime: delegation, coordination, state, and typed execution.
Memory layer: episodic memory, retrieval, compression, persistence, and replay.
Workflow engine: durable execution, retries, checkpoints, and human gates.
Evaluation systems: regression suites, trace review, cost tracking, and evidence.

Projects

emotional-memory — Research-grade memory layer for LLM systems. Python, PyPI, DOI 10.5281/zenodo.19640250, MIT.
orka — Agent orchestration in Rust. MCP/A2A protocols, multi-channel routing (Telegram/Discord/Slack/WhatsApp/HTTP), WASM plugins.
msg2agent — Agent-to-agent communication. X25519/Ed25519 E2E encryption, W3C DID identity, relay hub with P2P fallback. Go.
OpenFatture — Electronic invoicing (FatturaPA/SDI) + Lightning Network payments + local-first AI. Python, v1.1.0, MIT.

FAQ

How do you scope a new technical engagement?

I start with a written brief: problem, current stack, known failure modes, and success criteria. If scope is unclear, I propose a short discovery sprint with a working prototype and an eval plan.

How do you take LangGraph workflows to production and prevent regressions?

Production LangGraph needs durable state, explicit routing, typed tool failures, retries, and trace-level observability. I ship it with an eval harness that covers the failure modes that matter.

How do you build RAG systems that work in production?

I start with diagnostics, not a new vector database. Chunking, hybrid retrieval, re-ranking, metadata filters, and a frozen eval set usually matter more than changing embeddings.

Contact

info@gianlucamazza.it

github.com/gianlucamazza

linkedin.com/in/gianlucamazza