Verona Food Festivals
An Alexa skill that answers "what festivals are on this weekend?" for Verona and its province — data gathered every night from an events portal and queried by voice by date, town, area or theme.
Invocation: “Alexa, apri sagre veronesi”
The idea
From September to November the province of Verona is one dense calendar of sagre — village food festivals: chestnuts, wine, radicchio, lake fish. Finding out which one is on this weekend, in the next village over means opening a site, scrolling a list, filtering. I wanted to just ask for it: “Alexa, ask sagre veronesi what festivals are on this weekend.”
What it can do
Five kinds of question, all in natural Italian:
- today’s and this weekend’s;
- by date — “on 15 October”, “in November”;
- by town or area — “in Bussolengo”, “on Lake Garda on Saturday”;
- by theme or product — “chestnut festivals”, “wine ones in October”.
It reads four results at a time and asks “want to hear more?” for the rest. With no date, searches by place and theme look at the next 60 days. It recognises the province’s 98 towns, even in short form (“Peschiera”, “Villafranca”), and areas: Lake Garda, Monte Baldo, Valpolicella, Lessinia, Valdadige, East and Lower Verona.
How it’s built
A scheduled function runs every night on AWS Lambda, reads the portal’s map of Verona event pages and downloads only the ones that changed since last time (with a pause between requests, to go easy on the site). From each page it extracts the event’s structured data — name, dates, place — and saves it in a DynamoDB table. A second function answers Alexa: a few hundred rows are read in one go and filtered in memory, no fancy indexes needed. The infrastructure is described with AWS SAM, and the expected cost is negligible.
Alexa’s interaction model — the list of towns, areas, themes and phrases — isn’t hand-written: it’s generated from a couple of Python files, so adding a town or theme means editing a list and regenerating.
The pitfalls of real data
The first run on real events overturned more than one assumption:
- A place isn’t a town. 10% of events were left without a town: “Bonaldole, Zimella”, “S. GIOVANNI LUPATOTO”, “Lugagnano di Sona”, “Montorio” — hamlets, abbreviations, compound names. Matching now tries the individual parts of the place, expands abbreviations, looks for a town mentioned in the text, and uses a list of certain hamlets. Events in hamlets with no obvious town stay that way, but are announced by place name.
- Themes aren’t village names. Asking for “castagne” (chestnuts) mustn’t find Castagnaro: in theme search, town names don’t count.
- Alexa rejects some phrases. A sample phrase with a contracted article (“dell’{tema}”) was refused by the model and had to be reworded.
- Redoing everything: to reprocess every page after adding hamlets, the collection function accepts a “force” command that ignores the last-modified date.
A question of permission
The source portal has no terms of use, and its robots.txt forbids its internal API to
programs: so the skill uses only the public pages, never the forbidden API. For
personal use that might have been enough, but not for a public skill: so I asked for
permission to reuse the data, and it was granted.
Toward the Store
Then came the time to send it to Amazon: public distribution in Italian, the “event finder”
category, icons and publishing settings entered through the console. There’s a trap here:
those settings existed only in the console, not in the code, and a later ask deploy
would have wiped them. The manifest in the repository was therefore realigned with the
version actually submitted, and the rule is to re-download the package from the console
before any deploy after a hand edit. During review, moreover, the skill’s development
stage stayed locked: no deploys until Amazon replied.
Status
Published on the Alexa Store: submitted for certification on 2026-10-04, passed, and now available to the public in Italian. It’s tried with real data and covered by an automated test suite. The project is done; what’s left is maintenance and, if wanted, better coverage of not-yet-recognised hamlets. A practical note: a change to the Lambda goes straight to production, while a change to the interaction model or the manifest needs a new certification.
Stack
Python (ask-sdk) · AWS Lambda · DynamoDB · EventBridge · AWS SAM · Alexa Skills Kit