Emanuele Boscari

Designed and deployed PATBOT, a production RAG chatbot for public job competitions, reaching 5,000+ user interactions and reducing support requests by 30%. Built a multi-tenant agentic platform (Google ADK, LangGraph) on shared Kubernetes infrastructure serving 10 public administration clients. Created an LLM-as-a-judge evaluation framework achieving ~80% agreement with human annotators, reducing manual annotation work. Developed large alignment datasets and regression pipelines to re-verify behavior after prompt or feature changes, implemented robust ingestion (LLM-as-a-chunker) with support for tables and images, and added hierarchical retrieval with query revision. Performed red teaming with malicious prompts and ran an empirical study of 8 Italian summarization local small language models on H100 GPUs to select an on-premise model using both classical and LLM-based metrics. Built ML inference workflows and exposed outputs via rate-limited FastAPI services with k6 load testing and h

Designed architecture for registration and authentication of public Wi‑Fi access. Built real-time dashboards from Nutanix cluster data using React and Node.js, and contributed to turning open data into public-facing infographics about parking and public spaces.