Analyst · Data & AI Ethics @ Deloitte Advisory
ISO/IEC 42001 Lead Auditor · Cybersecurity Specialist

$ whoami
────────────────────────────────────────────────────────────────────────
 ROLE    Analyst · Data & AI Ethics @ Deloitte Advisory (Milan)
 WORK    agentic AI development: GenAI platform guardrails ·
         multi-agent codebase review · trustworthy AI · cybersecurity
 GOAL    LLM and agentic systems, secure by design
 CERTS   ISO/IEC 42001 Lead Auditor · Cybersecurity Specialist (EPICODE)
 EDU     MSc Astrophysics & Cosmology · University of Bologna
 NOW     AI Solutions Architect master @ ProfessionAI
         00-07 delivered · 08-10 planned
────────────────────────────────────────────────────────────────────────

My work at Deloitte has been shifting from assessing AI systems to building them. Today I develop agentic AI solutions on a proprietary GenAI platform, where I designed the risk methodology and the guardrails applied to every agent and app it ships, and I co-develop a multi-agent system that reviews agentic codebases by combining static analysis with LLM-based evaluation of governance, compliance and risk. The rest of my time goes to trustworthy-AI implementation (EU AI Act, ISO/IEC 42001, NIST AI RMF) and cybersecurity: risk assessments, security policies, access audits, awareness campaigns.

Astrophysicist by education (MSc, University of Bologna), I came to AI through ML research on gravitational lensing, with a thesis on automatic differentiation (PyTorch / TensorFlow). I co-authored HH4AI, a methodological framework for AI Human Rights Impact Assessment under the EU AI Act (LSAI 2025, Pisa). Off the keyboard, the mountains: ski mountaineering, alpinism, trail running.

Type-safe LLM vulnerability triage on Pydantic AI + a custom MCP server. Ten typed tools (NVD · OSV · CISA KEV · EPSS · Exploit-DB · MITRE ATT&CK · SBOM · Nmap) ground every answer into a validated TriageReport with a deterministic SSVC verdict. Build it, then threat-model it: MITRE ATLAS red-team battery, OWASP LLM Top 10 + ISO 42001 mapping, reproducible scorecard (resistance · retrieval MRR · cost / latency · calibration).

Live demo — keyless, replays real captured triages across the full SSVC ladder, with a scorecard.

One hands-on project per master module: RAG for a company knowledge base, two agentic capstones, multi-cloud architecture, ML deployment. Modules 00-07 delivered, 08-10 planned; full progress log in the repo README.

MSc thesis code grown into a PyTorch toolkit for parametric strong, weak and microlensing, extended with deep-learning and LLM-assisted case studies.

agentic          LangGraph · LangChain · Pydantic AI · MCP · n8n
llm / rag        OpenAI / Azure / Anthropic APIs · Hugging Face Transformers
                 sentence-transformers · LiteLLM · ChromaDB · FAISS · BM25 hybrid · Ollama
engineering      Python · FastAPI · Flask · Docker · pytest · Git · GitHub Actions · Bash · SQL
ai-assisted dev  Claude Code (daily driver) · GitHub Copilot · agentic coding workflows
cloud / mlops    AWS SageMaker · Azure ML / OpenAI · Vertex AI (GCP)
ml / data        PyTorch · TensorFlow · scikit-learn · pandas · NumPy
ai security      OWASP LLM Top 10 · MITRE ATLAS · LLM red teaming · threat modeling
governance       EU AI Act · ISO/IEC 42001 · NIST AI RMF
vapt             Nmap · Metasploit · Burp Suite · SQLmap · Nessus · Wireshark · Splunk · Semgrep

Full CV, experience and education on LinkedIn.