IT / Technical Project Manager · PMO, Governance & AI Automation · Torino
Fabio Ferrari.
Project Manager — governo l'enterprise e orchestro i sistemi AIProject Manager — governing the enterprise, orchestrating AI systems
Delivery e governance PMO enterprise per CNH Industrial,Porini,Digital Thinks — 15+ commesse IT, un team di 7 PM in matrice, €1.8M+ di risparmi negoziati. Da marzo 2025 progetto, costruisco e opero 10+ sistemi AI end-to-end, in produzione 24/7.Enterprise PMO delivery and governance for CNH Industrial,Porini,Digital Thinks — 15+ IT engagements, a 7-PM team in a matrix, €1.8M+ in negotiated savings. Since March 2025 I design, build and operate 10+ end-to-end AI systems, in 24/7 production.
Governance applicata all'AI: registro di claim verificati anti-allucinazione, quality gate umano, purge documentato dei contenuti fabbricatiGovernance applied to AI: anti-hallucination registry of verified claims, human quality gates, documented purge of fabricated content
FinOps applicato: cost tracking multi-provider cloud, budget cap per servizio, placement dei workload guidato dai costiApplied FinOps: multi-provider cloud cost tracking, per-service budget caps, cost-driven workload placement
10+ sistemi progettati, costruiti e operati in produzione 24/710+ systems designed, built and operated in 24/7 production
LLM orchestration, RAG, data pipeline, observability — sistemi veri in produzioneLLM orchestration, RAG, data pipelines, observability — real systems in production
Effort di sviluppo stimati e contestati con cognizione di causaDev effort estimated and challenged with first-hand knowledge
AI & Automation Consultant — Progetti Indipendenti
Mar 2025 — oggipresent
Torino · Scelta strategica: colmare il gap tra governance e delivery tecnica.Turin · Strategic choice: closing the gap between governance and technical delivery.
Progettato e messo in produzione un portfolio di 10+ sistemi AI/automation end-to-end su un backbone architetturale condiviso, gestendo in autonomia l'intero ciclo: discovery, architettura, sviluppo, deploy, observability ed evoluzione.Designed and shipped a portfolio of 10+ end-to-end AI/automation systems on a shared architectural backbone, owning the full lifecycle solo: discovery, architecture, development, deployment, observability, evolution.
Gestita infrastruttura self-hosted in produzione 24/7: 25+ container Docker, VPN mesh, monitoring centralizzato, backup automatizzati, CI/CD — stessi pattern enterprise (alta disponibilità, telemetry, security by design), a scala personale/small-team.Operate self-hosted 24/7 production infrastructure: 25+ Docker containers, VPN mesh, centralized monitoring, automated backups, CI/CD — same enterprise patterns (HA, telemetry, security by design), at personal/small-team scale.
Un engagement commerciale consegnato a un cliente esterno (startup torinese): piattaforma di lead generation B2B con CRM, scoring e go-to-market.One commercial engagement delivered to an external client (Turin startup): B2B lead-generation platform with CRM, scoring and go-to-market.
Comprensione diretta di come l'AI trasforma l'SDLC e la stima degli effort — uso quotidiano di coding agent (Claude Code, Cursor, Antigravity), prompt engineering, pattern di orchestrazione.First-hand understanding of how AI reshapes the SDLC and effort estimation — daily use of coding agents (Claude Code, Cursor, Antigravity), prompt engineering, orchestration patterns.
Project Manager | PMO Lead — DGS S.p.A.
Lug 2021 — Mar 2025
Italia (hybrid) · Delivery e governance su tre clienti enterprise.Italy (hybrid) · Delivery and governance across three enterprise clients.
Digital Thinks— Digital Transformation· Ott 2024 – Mar 2025
Ridisegno della pipeline di erogazione (commesse, fatturazione, contrattualistica) e gestione del SAL settimanale (stato avanzamento lavori) sulle commesse formative.Redesigned the delivery pipeline (engagements, billing, contracts) and ran the weekly work-progress review (SAL) across training engagements.
Throughput formativo più che triplicato (da ~15 a 50+ corsi/mese), con dashboard Power BI real-time sui KPI critici e coordinamento di 5 fornitori tecnici.Training throughput more than tripled (from ~15 to 50+ courses/month), with real-time Power BI dashboards on critical KPIs and coordination of 5 technical vendors.
Porini— PMO & Multi-Client Governance· Lug 2022 – Ott 2024
PMO end-to-end su 15+ commesse IT simultanee, inclusi progetti di migrazione cloud (Azure, AWS): forecast accuracy >95%, -40% escalation.Ran end-to-end PMO across 15+ concurrent IT engagements, including cloud migration projects (Azure, AWS): forecast accuracy >95%, -40% escalations.
Coordinato in matrice un team di 7 Project Manager (assignment, prioritizzazione, quality-gate, escalation — no riporto gerarchico).Coordinated a 7-Project-Manager team in matrix (assignment, prioritization, quality gates, escalation — no line management).
Governance automatizzata (Jira custom + Power BI live + template Confluence): reporting da 8 a 2 ore/settimana, -55% effort.Automated governance (custom Jira + live Power BI + Confluence templates): reporting from 8 to 2 hrs/week, -55% effort.
Ottimizzato un portfolio licenze software da €5.2M/anno con analisi di utilizzo, negoziazione e riallocazione costi cross-region: -37% costi, €1.8M+ saving sul ciclo biennale.Optimized a €5.2M/year software licensing portfolio via usage analysis, vendor negotiation and cross-region cost reallocation: -37% cost, €1.8M+ saved over the two-year cycle.
Gestione dell'infrastruttura server virtualizzata (ambienti R&D, telemetria macchinari) e del lifecycle/obsolescenze del parco workstation e server.Managed the virtualized server infrastructure (R&D environments, machinery telemetry) and the lifecycle/obsolescence of the workstation and server estate.
Change management e adoption tool su team Italia–USA–India: 100% adoption in 8 settimane. Coordinato l'avvio dei poli produttivi di Racine (USA) e New Delhi (India).Change management and tool adoption across Italy–USA–India teams: 100% adoption in 8 weeks. Coordinated the launch of the Racine (USA) and New Delhi (India) production hubs.
Consulente FinanziarioFinancial Advisor— ING Italia
Apr 2019 — Giu 2021
Gestione portafoglio clienti, analisi profili di rischio e pianificazione patrimoniale in contesto regolamentato.Client portfolio management, risk-profile analysis and wealth planning in a regulated context.
03// Sistemi in produzione// Production systems
Sistemi in produzioneProduction systems
Il differenziatore: sistemi reali, costruiti e operati in prima persona, con metriche verificabili sul codice.The differentiator: real systems, built and operated first-hand, with metrics verifiable on the code.
Produzione 24/7 su infrastruttura self-hosted (cluster ARM64), scala personale/small-team. Codice privato — architettura, demo e walkthrough su richiesta.24/7 production on self-hosted infrastructure (ARM64 cluster), personal/small-team scale. Private code — architecture, demos and walkthroughs on request.
in produzione dal 2025 · oggi v2in production since 2025 · now v2~/systems/ai-gateway
v1, costruita da zero (2025–26): gateway self-hosted OpenAI-compatible con routing adattivo (Thompson sampling) su 13 provider e 163 deployment di modelli, hedging della tail-latency, rate limiting GCRA, circuit breaker a 3 stati con auto-heal, key-pool cooldown-aware, 15+ metriche Prometheus, OpenTelemetry, SLO burn-rate multi-window. v2, in produzione oggi: quando il problema vero si è spostato dal routing intelligente all'aggregazione e rotazione di quote multi-provider, ho ritirato la v1 e ricostruito su un motore open-source maturo — failover, policy di quota e di costo (budget cap per provider) e hardening restano progettati e operati da me.v1, built from scratch (2025–26): self-hosted OpenAI-compatible gateway with adaptive routing (Thompson sampling) across 13 providers and 163 model deployments, tail-latency hedging, GCRA rate limiting, 3-state circuit breaker with auto-heal, cooldown-aware key-pool rotation, 15+ Prometheus metrics, OpenTelemetry, multi-window SLO burn-rate. v2, in production today: once the real problem shifted from intelligent routing to aggregating and rotating multi-provider quotas, I retired v1 and rebuilt on a mature open-source engine — failover, quota and cost policy (per-provider budget caps) and hardening remain designed and operated by me.
13
provider
163
model deployment
235
file di testtest files
90%+
success rate
ARM64
edge hardware
decisioni & tradeoffdecisions & tradeoffs
Thompson sampling al posto di una classifica statica dei provider: i free-tier degradano in modo imprevedibile — la qualità reale va appresa in continuo (explore/exploit), non scritta a mano.Thompson sampling instead of a static provider ranking: free tiers degrade unpredictably — real quality has to be learned continuously (explore/exploit), not hand-ranked.
Rate limiting GCRA invece di finestre fisse: flusso levigato, niente raffiche al reset della finestra.GCRA rate limiting instead of fixed windows: smooth flow, no thundering herd at window reset.
Il costo vero del build non è scriverlo, è mantenerlo: la v1 valeva quel costo finché il routing era il differenziatore. Quando ha smesso di esserlo, il costo è rimasto e il valore no — da lì la v2.The real cost of building isn't writing the code, it's maintaining it: v1 was worth that cost while routing was the differentiator. When it stopped being one, the cost stayed and the value didn't — hence v2.
Personal AI Knowledge SystemPersonal AI Knowledge System
R&D
"Secondo cervello": knowledge graph temporale + retrieval ibrido a 3 lane (denso + BM25 + graph walk, fusione RRF) con variante HippoRAG-2 su Personalized PageRank. Agente proattivo con confidence-routing e critic anti-contraddizione."Second brain": temporal knowledge graph + 3-lane hybrid retrieval (dense + BM25 + graph walk, RRF fusion) with a HippoRAG-2 variant on Personalized PageRank. Proactive agent with confidence-routing and an anti-contradiction critic.
decisioni & tradeoffdecisions & tradeoffs
Fusione RRF invece di un reranker addestrato: niente training data, comportamento spiegabile, ogni lane degrada in isolamento.RRF fusion instead of a trained reranker: no training data needed, explainable behaviour, each lane degrades in isolation.
L'eval A/B ha smontato il mio stesso numero: recall@10 0.55 era un artefatto di un corpus con ~45% di duplicati (corretto: 0.85). Prima l'igiene dei dati, poi le feature.The A/B eval broke my own number: recall@10 of 0.55 was an artefact of a corpus with ~45% duplicates (corrected: 0.85). Data hygiene first, features second.
Critic anti-contraddizione + confidence-routing: sotto soglia il sistema tace — meglio nessuna risposta che una allucinata.Anti-contradiction critic + confidence routing: below threshold the system stays silent — no answer beats a hallucinated one.
1.394 test5 output channelRAG · KG
Lead-Gen B2B per startupB2B Lead-Gen for a startup
startup TorinoTurin startup
Engagement commerciale (startup torinese di installazioni immersive): scraping multi-fonte, enrichment contatti, scoring deterministico a 100 punti, CRM self-hosted con pipeline Kanban, outreach + go-to-market a 90 giorni.Commercial engagement (Turin immersive-installation startup): multi-source scraping, contact enrichment, deterministic 100-point scoring, self-hosted CRM with Kanban pipeline, outreach + 90-day go-to-market.
Piattaforma full-stack di finanza personale: 9 engine analitici schedulati (forecast, anomaly, Monte Carlo sulla probabilità di rovina), 4 feature LLM con circuit breaker (chat text-to-SQL sandboxed), security rewrite documentata (WebAuthn, CSRF, CSP).Full-stack personal-finance platform: 9 scheduled analytics engines (forecast, anomaly, Monte Carlo ruin probability), 4 LLM features behind a circuit breaker (sandboxed text-to-SQL chat), documented security rewrite (WebAuthn, CSRF, CSP).
decisioni & tradeoffdecisions & tradeoffs
Feature LLM dietro circuit breaker: la piattaforma resta pienamente utilizzabile con l'LLM giù — degradazione pianificata, non best-effort.LLM features behind a circuit breaker: the platform stays fully usable with the LLM down — planned degradation, not best-effort.
Chat text-to-SQL sandboxed: l'LLM non tocca il database direttamente — query validate in sandbox. La capability senza la superficie d'attacco.Sandboxed text-to-SQL chat: the LLM never touches the database directly — queries go through validation in a sandbox. The capability without the attack surface.
Security rewrite (WebAuthn, CSRF, CSP) su un'app già in uso: retrofit trattato come progetto — scope e verifiche documentate — non come patch.Security rewrite (WebAuthn, CSRF, CSP) on an app already in use: retrofit run as a project — documented scope and verification — not as a patch.
Servizio condiviso di browser-automation engine-pluggable: 6 backend di rendering dietro un'unica API con fallback a caldo, motore anti-bot YAML-driven (8 segnali, 31 predicati), sessioni "human-shaped" — consumato da 3+ progetti.Shared engine-pluggable browser-automation service: 6 rendering backends behind one API with hot fallback, YAML-driven anti-bot engine (8 signals, 31 predicates), human-shaped sessions — consumed by 3+ projects.
313 test6 engineinfra condivisa
IoT & Device FleetIoT & Device Fleet
produzioneproduction
Flotta di 60+ device IoT gestita end-to-end con 100+ automazioni event-driven: telemetria continua, health-check con rollback automatico sugli update, deploy su edge (single-board computer ARM64); hub di geolocation & asset-tracking GPS/BLE con 14 job schedulati.Fleet of 60+ IoT devices managed end-to-end with 100+ event-driven automations: continuous telemetry, health checks with automatic rollback on updates, edge deployment (ARM64 single-board computers); GPS/BLE geolocation & asset-tracking hub with 14 scheduled jobs.
decisioni & tradeoffdecisions & tradeoffs
Update della flotta dietro health-gate con rollback automatico: nessun update manuale, ma nessun update cieco — se un servizio non torna sano, torna da solo alla versione precedente.Fleet updates behind a health gate with automatic rollback: no manual updates, but no blind updates either — if a service doesn't come back healthy, it rolls back on its own.
Scala personale dichiarata, ciclo di vita aziendale: provisioning, telemetria, update sicuri e incident response sono gli stessi di una flotta enterprise — cambia il numero di device, non il metodo.Personal scale, declared openly — enterprise lifecycle: provisioning, telemetry, safe updates and incident response are the same as in an enterprise fleet; the device count changes, not the method.
Geolocation self-hosted: i dati di posizione non lasciano l'infrastruttura — la capability dell'asset-tracking senza cedere dati a un cloud terzo.Self-hosted geolocation: location data never leaves the infrastructure — asset-tracking capability without handing data to a third-party cloud.
Personal AI Knowledge System — retrieval ibrido + brainPersonal AI Knowledge System — hybrid retrieval + brain+
Tre lane di retrieval fuse via RRF, poi un brain con confidence-routing e critic anti-contraddizione verso 5 canali di output.Three retrieval lanes fused via RRF, then a brain with confidence-routing and an anti-contradiction critic feeding 5 output channels.
Collector isolati sul nodo egress, consegna crash-safe via WAL al nodo always-on, analisi LLM e hand-off al knowledge graph.Collectors isolated on the egress node, crash-safe WAL delivery to the always-on node, LLM analysis and hand-off into the knowledge graph.
Resale-Intelligence Platform — qualificazione a convergenzaResale-Intelligence Platform — convergence qualification+
Estrazione distribuita → filtro → qualificazione vision multi-provider con failover → pricing → scorer a convergenza di 6 segnali.Distributed extraction → filtering → multi-provider vision qualification with failover → pricing → 6-signal convergence scorer.
05// Competenze// Skills
CompetenzeSkills
Dove governance e delivery tecnica si incontrano.Where governance and technical delivery meet.
