Ugo Lattanzi
This page is also available inItalianoDeutsch
Work

Nine systems, and what moved.

Client names are withheld deliberately. What matters is the pattern and the outcome, and both are here in full. The last three predate generative AI: they are the platforms where I learned what actually holds up in production.

Guidelines · research body

Automated the production and update of sector guidelines with an agentic loop on the Claude API over a RAG layer of the guideline corpus, replacing a manual desktop workflow. A process that took six months now produces an updated or new guideline in hours. Validation was built with the body's own reviewers over a long iterative programme whose feedback continuously refined the pipeline and its quality metrics.

Six months → hours
The full account: Six months to hours →

Event & member operations · same body

Designed and currently building a multi-agent orchestrator with human-in-the-loop checkpoints across event operations — attendee support, ticketing, registration, admissions. The guideline platform opened the door; the engagement then expanded into the organisation's operational core.

In build

Audience analytics · media group

Put terabytes of smart-TV viewing data from markets worldwide in the hands of business users, replacing Grafana dashboards a specialist had to prepare in advance with a conversational interface that answers in natural language and generates charts and reports on demand. Databricks for TB-scale transformation, Fabric ETL into MongoDB, retrieval via function calling.

A specialist → a plain question
The full account: How to test an LLM system →

Customer support · utility company

Assistants over a RAG knowledge base: the system connects to the support mailbox, drafts a reply to every request, and the operator stays the last step — rating, editing, sending — with ratings feeding an improvement dashboard. Continuous learning was deliberately left out because the customer needed the knowledge base under their own control, and the design was built around that constraint rather than arguing it away.

Assistive by design
The full account: When not to automate →

Clinical decision support · public hospital group

Worked alongside the hospital's own technical team to bring AI in-house, building support procedures for physicians during consultations — cross-referencing patient history and enriching the live visit through audio capture, with privacy controls throughout. As much customer enablement as build.

Built with their team

Quote automation · insurance

Automated the request, comparison and evaluation of quotes across the major carriers in the market, and the guidance of the customer through the result.

Hours per case → minutes

Parts trading and returns · multinational automotive group

Designed, and partly built, the parts trading and returns platform for the group's entire European network: orders, stock and availability across warehouses, price lists and discount tiers by customer class, returns with time windows and penalties, invoicing and shipment tracking. Several ageing, slow systems replaced by a single microservice platform — React on the front, MongoDB, Kafka for events, Elasticsearch for search across catalogues of millions of part numbers, Auth0 for identity — on Kubernetes in AWS. Operations that used to cross several applications now sit in one flow.

Several ageing systems → one platform

Used-vehicle remarketing · premium automotive brand

Designed and built with my team the used-vehicle trading system — young used car and off-lease — for the Italian arm of a premium brand: full auctions with reserve price, automatic proxy bidding, closing extensions and allocation, alongside a direct sales channel. Cars used to be traded over the phone, one at a time and few at a time; the platform moved that to tens of thousands of vehicles sold online, to the point where the parent company took an interest in the model.

Phone calls → tens of thousands of cars online

Cloud platform · international fitness equipment manufacturer

Technical lead on the cloud platform that ties equipment, apps and web together. The monolith was re-architected into more than fifty microservices on AWS ECS Fargate, fully event-driven with SNS and SQS, with a 10 TB sharded MongoDB cluster and gRPC for server-to-server communication. Alongside the architecture I defined the cloud strategy across the three touchpoints — mobile, equipment, web — with automated provisioning, security and CI/CD, and I drove R&D and enablement: new technologies and agile practices brought in, with mentoring for teams across business units.

Monolith → 50+ microservices

If this is the problem on your desk, write to me.

info@ugolattanzi.com
Ugo Lattanzi — Applied AI Architect No cookies. Aggregate, cookieless analytics via Cloudflare.