Ontology engineering, knowledge graphs and governed LLM agents

Machine learning in production

Enel Group · 2017–2023

In a six-person Agile team, I co-built and maintained for six years a machine learning platform in production on AWS for predictive maintenance and investment targeting, from requirements to production, adoption and ongoing operation. I selected Databricks, personally developed the platform’s entire ETL layer and did the modelling: feature selection, XGBoost training and evaluation, and thresholds for operational use.

Result: the platform contributed to reducing average customer interruption minutes by up to 30% in some territories (internal before-and-after measure).

Platforms in production, managed as CTO

Sistemi & Automazione · June 2026 – September 2026

I managed two platforms in production: a graph-analysis platform with a multi-agent tool that users could query about the graph data, and a due-diligence platform with about 60 connectors to different data sources. I handled their production issues end to end.

Boundary: both platforms existed before I joined. I managed them; I did not build them.

OCOR

A sovereign, vendor-neutral semantic core that connects ontology definitions, evidence, governed decisions and controlled actions. I designed its ontology-as-code metamodel and built its ontology compiler: it type-checks YAML or JSON ontology sources, rejects unresolved or cyclic references, and turns each release into a deterministic, content-addressed canonical model with a semantic diff and a compatibility class. A breaking release is rejected without a migration plan. AI agents use only typed, policy-filtered named queries and typed tools, with budgets and a kill switch.

Limitation: an implementation in development on a wholly synthetic scenario, not a production system. Qualification is in progress.

View OCOR repository (opens in a new tab)

Enterprise Code Intelligence

Graph-RAG over a codebase, with event-driven updates and vector, full-text and graph retrieval. Services in Python, Go and Rust; Neo4j, Qdrant, OpenSearch, Kafka and Debezium, PostgreSQL, Redis.

Limitation: a runnable proof of concept, not a production service.

View Enterprise Code Intelligence repository (opens in a new tab)

Enterprise Sovereign AI Decision Matrix

A governed AI workflow that routes the model call to a provider by data-sovereignty rules, requires human approval, records every decision and tracks latency, token use and cost. Next.js, FastAPI, LangGraph, PostgreSQL.

Limitation: a fixed, typed workflow with governed model calls, demonstrated on synthetic data with one real routed model call. It is not a general-purpose multi-agent runtime or a multi-tenant production platform.

View Decision Matrix repository (opens in a new tab) · Open live demonstration (opens in a new tab)

IADF Platform

The core of a policy-governed, AI-assisted software factory: contract-first schemas, an explicit state machine, PostgreSQL leasing, model routing and fail-closed controls. Its four-role agent engine (architect, implementer, adversarial reviewer, diagnostician) takes coding agents from specification to tested code.

Limitation: an inspectable core and architecture proof of concept, not a complete production software factory.

View IADF Platform repository (opens in a new tab)

Agent demonstrations

In the HR onboarding orchestrator, an agent runs an onboarding workflow across simulated systems through tool calls. In the hiring panel and the innovation board, six agents assess a case from different perspectives. In all three, code validates the agents’ output and a person decides.

Limitation: demonstrations on simulated data, not production systems.

Explore the demonstrations

Can LLM Agents Care About the World?

A formal control framework (2026) in which an LLM agent judges improvement or deterioration from an external, auditable indicator, with public simulation code.

Limitation: a theoretical preprint, not peer reviewed.

Read about the preprint · ResearchGate, DOI 10.13140/RG.2.2.25820.55685 (opens in a new tab)

Method

Core Question → Literature Review → Stated Assumptions → Success & Failure Criteria → Instrumentation → Validation.

I separate evidence, inference and hypothesis, and I set evaluation criteria before I claim results. I instrument systems to monitor quality, cost, latency and failure rates, and I revise architectural choices against observed results and known limitations.

Applied governance

As CTO, I authored the company AI policy (EU AI Act, GDPR and Italian Law 132/2025, prepared with a view to ISO/IEC 42001 certification): risk-based classification of AI use, human oversight and graduated autonomy levels for AI agents. I also authored the guidelines for AI-assisted software development, with controls for requirements, testing, traceability and validation.

Boundary: these are internal company documents, not a public standard or an externally audited framework.

Chief Technology Officer & Head of Research and Innovation

Sistemi & Automazione S.p.A. · June 2026 – September 2026 · Rome, Italy

I established the technology function, with architecture and research units: 5 direct and 15 functional reports. I assessed the organisation, products and engineering practices in two weeks, proposed a new decision-making process, agreed it with business-area heads and Tech Leads, and defined governance rules for projects and technology choices (RACI).

Enel Group

E-Distribuzione S.p.A. · May 2008 – June 2026

I created the unit that trains a technical organisation of 1,500 people; at the end of 2025 it ranked first of the 12 training units in Italy for training hours per person. I monitored Grid Blue Sky, Enel Grids’ global digitalisation programme, and reported its progress to the CEO. I led a six-month cross-functional team that turned emerging technologies into roadmaps and investment choices.