Understanding people, reading data, building with AI

The hard parts

Three layouts in thirteen years

I wrote an extraction step that recognises headlines, columns, boxes and recurring sections, so each article stays whole.

Two kinds of question

“When was the church of the Madonna di Juso built?” is about meaning; “What came out in May 2022?” asks for an exact fact. The first needs embeddings, the second needs dedicated tools.

Trust, in a small town

Answers are structured: every statement is tied to a specific article.

Three choices and their trade-offs

• Hybrid search, not vector search alone. More parts to maintain, but numbers and dates come from the data, not from a guess.
• Show only the sources actually cited. The list is shorter, but every link backs up a sentence in the answer.
• A screen that guides the first question. One more step, but it helps people ask what the archive can answer.

The bot remembers the conversation and can show the cover of any issue. It is a volunteer project, built for a small newsroom and a town with a lot to tell.

• Tracking. Tag Manager for UX events, advertising and new features; GA4 with funnels, key events and channels consistent with UTM tags; consent is handled by the CMP.
• History. Python scripts save Search Console and Analytics data on the company’s servers; a move to Metabase was under way.
• Content and search. Embeddings and cosine similarity to measure how relevant a page is compared with competitors; in 2025 a backlink audit shifted the focus from spam to opportunities.

Three choices and their trade-offs

• Start from the questions, not from ready-made reports. Fewer charts, but each one answers a question someone in the company actually asks.
• Keep our own data history. The scripts need maintaining, but trends since 2012 survive when the platforms delete old data.
• Align GA4 channels with UTM tags. Extra work up front, but campaign reports match what marketing had planned.

I worked as a UX designer alongside designers, developers and marketers. Our users read in a hurry, jumping from one paragraph to the next and stopping on strong images. A seven-stage customer journey map showed, for each stage, what we already did and what we could do.

A/B tests went live with Google Experiments, now discontinued: a switch instead of a button, two versions of the information request form, different sign-up flows. The numbers stayed with the company; the versions that worked better are the ones that stayed live.

The Archiportale brief, in one questionHow can we get construction, architecture and design professionals to read as many news stories as possible on archiportale.com, and so justify higher advertising rates?

• Competitor analysis. 19 magazines grouped into three patterns: asymmetric inverted pyramid, symmetric inverted pyramid, symmetric columns.
• Advertising. On mobile, full-screen skins disappear and vertical banners become square.
• From wireframes to code. HTML pages on 12- and 16-column grids with Foundation 6, files versioned in TFS, and AMP under evaluation for mobile.

Three choices and their trade-offs

• Keep every URL and the paths to Projects, Designers and Products. Less freedom to redesign the structure, but traffic from Google and from the rest of the network was protected.
• Gather the news in one section, with the topics always in view. One more level, but every topic stays one click away from any page.
• Test small changes on live traffic. Slower than one big redesign, but every change rests on what users actually did.

The answers told us to keep the switch. We made three small fixes instead: more contrast between the on and off states, a larger click area, and clearer links to the categories with products on sale.

The most useful detail came from two developers who did not know the product: one spent a long time looking for the category in the drop-down menus, the other could not find the switch and kept using the price filter.

• The task. Find only the products on sale in the TV units category: between 14 and 56 seconds.
• How they did it. 7 with the search engine, 4 with the drop-down menu, 1 by browsing the categories, 3 by switching approach halfway.
• Short or long label. “In vendita” (on sale) or “Mostra solo prodotti in vendita” (show only products on sale): 8 to 7.

• Semantic HTML. Main categories are now H2s and subcategories bold text; each category has its own description, for a better text-to-code ratio.
• Structured data. Products, brands, breadcrumbs and news described according to schema.org.
• On a category page the mobile speed score went from 69 to 75.