I turn messy data into systems that decide
6+ years at Elmec Informatica, from Service Desk to technical reference for the AI team. I build production LLM systems and set the standards the team builds on
LLM Gateway
OpenAI-compatible LLM gateway in Go: one API in front of OpenAI, Anthropic, and on-prem models (vLLM, Ollama), with token-level access control, RBAC, budgets, rate limiting, and OpenTelemetry observability, plus a 40-tool MCP admin server. In production at Elmec.
LLM Availability Dashboard
Real-time monitoring for LLMs across our Kubernetes infrastructure (kServe, vLLM): a live probe loop tracks model health, and when one drops an Agno agent auto-diagnoses the cause, inspecting pods, events, and logs through read-only tools and testing the model endpoint directly, with its reasoning streamed live to the UI.
Phishing Email Triage
FastAPI service that parses incoming emails and scores phishing risk with an LLM, wired into the cybersecurity team’s ticketing flow. Langfuse-managed prompts, deployed on Kubernetes, and the basis for my public AI & Security talks.
AI Documentation Assistant
LLM workflows for technical authors: validates draft sections against guidelines and offers suggest / rework / extract actions, streamed token-by-token. Structured output via Instructor, prompts in Langfuse, deployed on Kubernetes.
Ticket Classifier
Classical-ML service that routes support tickets by type and forwarding group across 50+ enterprise clients: TF-IDF + SVM over FastAPI, predictions persisted to PostgreSQL. ~1,900 interactions/month.
Generative UI Dashboard
Upload a CSV / Excel / JSON file and get an interactive dashboard: one LLM call maps columns to chart specs, then everything renders deterministically through ECharts templates, with no generated code. Drag-and-drop layout, SSE streaming. Earlier iterations explored agentic generation.
Procedure Knowledge Base
A knowledge base for retail IT support that routes questions to verified procedures instead of rewriting them: a vision LLM enriches screenshots and diagrams, an LLM builds a routing index, and a query API answers with citations to the source docs.
Enterprise Site Search
A search engine on OpenSearch for a corporate site: hybrid full-text search with query optimization, score-based re-ranking, real-time autocomplete, and a Trino to OpenSearch ingestion pipeline, plus continuous LLM-as-a-Judge evaluation of result quality. Built with the team.
Conversational AI
Graph-orchestrated support chatbot (semantic RAG retrieval, LLM-as-judge escalation, and multi-turn info collection) that I contributed to with the Elmec AI team.
This Site
The site you’re reading. Next.js + React with a generative hero (a feed-forward neural network in a particle field that resolves my name out of noise) and a role-semantic OKLCH color system. Designed and built end-to-end.
Generative AI & LLM
Where I spend most of my time. Agentic systems with tool-calling, from an Agno agent that diagnoses live model outages to graph-orchestrated chat, backed by RAG, prompt engineering, and LLM-as-a-Judge for continuous quality. The basis for 6 talks (2024–2026, 2 external incl. ATED).
LLM Providers & Deployment
Hybrid cloud / on-prem deployment strategy across providers. Currently deepening the end-to-end on-prem lifecycle with focus on vLLM (serving, monitoring, scaling, version management).
MLOps & Observability
I introduced Langfuse as the company-wide observability baseline, now adopted across every internal AI project for tracing, prompt versioning, evaluation, and cost monitoring.
Vector Databases & Search
Semantic and hybrid search, re-ranking, continuous validation via LLM-as-a-Judge, powering RAG and enterprise search systems. Qdrant-certified (01/2026).
Frameworks & Programming
Production stack across Python services, ML pipelines, and lightweight UIs. React for full-stack work; FastAPI / Flask / Gradio depending on the surface.
Data Science & Visualization
NLP and predictive modeling. Automated EDA tooling and dashboard work directly support the Generative UI Dashboard build.
Cloud, DevOps & Infrastructure
Foundations from the System Engineering era; still actively shaping AI deployment decisions across cloud and on-prem environments.
L1/L2 incident and escalation management with ITSM: the foundation of technical analysis and communication.
Stabilized Windows / Linux, cloud and virtualization; automated ops with Python / PowerShell and pioneered the first ML on ticketing.
Technical reference for the AI team on LLM, RAG and MLOps; owns the hybrid deployment strategy and observability standard.
Data Science & AI
2020 – PresentSelf-Study
- Specialist courses on DeepLearning.AI, Coursera, O'Reilly: ML, DL, NLP, MLOps, LLMs, agentic systems
- External validation via vendor-neutral certifications (NVIDIA, Qdrant, HuggingFace)
B.Sc. Computer Science (interrupted)
2018 – 2022Università degli Studi di Milano
- Path interrupted by deliberate choice to specialize in Data Science & AI through targeted training and enterprise application; the 2020-onwards career trajectory confirms the coherence of that choice.
- Foundation acquired: programming, algorithms, databases, software engineering, statistics
Scientific High School Diploma
2013 – 2018Liceo Scientifico Statale Arturo Tosi
- Grade: 78/100