Behavioral evaluation framework for agentic LLM systems
Task-oriented evaluator with multi-rubric LLM-as-a-judge and a 200+-scenario multi-turn adversarial probe suite for a national-scale PA chatbot.
Projects
Systems I’ve built and shipped to production. The CSI Piemonte projects are described at the level of detail already public in the linked articles; all my personal projects, on the other hand, are open source and fully inspectable on my GitHub.
Two-stage search engine for public tenders
Camilla’s retrieval engine: 101K production queries, 2.3s average end-to-end latency, 97% pass rate on a several-hundred-query regression suite.
Internal MCP server for agent tool composition
Reusable primitives and declarative YAML composition: building an agent for a new client went from 4-5 days to half a day.
AI Voice Assistant
Open-source phone agent that answers calls on your behalf: two interchangeable conversation engines, handcrafted turn taking, under a second to first audio.