I built Weather Alerts Spain after a red wind warning caught my family by surprise while we were hiking. I wanted a reliable way to be notified of weather warnings in English as I am not tuned into local Spanish media.
The result is a live email service that lets subscribers choose the areas and warnings which are delivered to their email inbox. Red alerts are delivered immediately while orange and yellow warnings are sent in scheduled summaries up to three times a day. It was my first product built with AI coding tools, starting in Lovable before I moved it to Cloudflare using Claude Code.
Product screens and details
Sign up
Sign up with one field and one promise: official AEMET warnings, delivered in English.
Choose alerts
Pick your region and areas, the severity levels you want to hear about, and when your summaries arrive.
Confirmed
See exactly what you will receive, and when.
Your emails
Red alerts go out straight away. Summaries of orange and yellow warnings arrive when you choose.
Running it
Is everything working? One screen for the owner, using sample data.
Select a screen to enlarge it. The admin screen uses sample data.
The story behind the project
The product
Weather Alerts Spain is an automated email notification service built around official weather warning data from AEMET, Spain's State Meteorological Agency. A scheduled Cloudflare Worker checks AEMET's national ATOM feed every five minutes. The feed contains links to individual Common Alerting Protocol (CAP) XML documents, each describing a warning affecting one or more geographical zones.
The application retrieves these documents and parses them into structured records. It extracts the warning type, severity, affected geographical zones, start and end times, updates or cancellations. The geographical codes are matched against reference data covering Spain's regions and warning areas, while weather phenomena are mapped to English descriptions. The records are then normalised and stored in a Cloudflare D1 database. The system identifies duplicate warnings and maintains the history of updates to existing alerts, rather than treating every feed entry as an entirely new warning.
Subscribers register using their email address and select the areas they want to monitor, their preferred warning levels and when they want to receive summaries. The application compares stored warnings against these preferences, including the exact geographical areas affected. Scheduled processing then compiles eligible warnings into email summaries, with up to three delivery windows each day. Red warnings have separate delivery rules: active or imminent warnings are prioritised for immediate delivery, while warnings further in the future can be included in a scheduled summary if it will arrive before the warning begins.
Emails are generated from the stored warning information and sent through SendPulse. Where AEMET publishes separate zone-level records for the same underlying red warning, the service can combine them into a single notification for the subscriber. The system also tracks delivery attempts and failures, and uses automated checks to monitor whether data ingestion and email processing are running. All warning classification and routing decisions use defined application rules rather than asking an AI model to interpret the risk.
Building it with AI
I selected a relatively simple user-facing product that still required authentication, external data, persistent preferences and automated email delivery. I prepared the requirements with ChatGPT, Claude and Gemini, divided the initial work into thirteen tickets and built the first version in Lovable and Supabase. I subsequently migrated the application to Cloudflare using Claude Code to manage the rebuild. The decision was driven largely by cost: I found Lovable's credit-based development model expensive as the project became more complex, while Cloudflare offered a more economical hosting solution.
I treated it as a product exercise rather than a single prompt to generate an app. The first specification covered the essential user journeys, the database, sign-in, delivery rules and what should happen when the external data changed. Breaking the work down before opening Lovable made it easier to see which parts were genuinely complete and which still needed verification.
The geography issue was a good example of how quickly real data can invalidate a neat specification. My original regional model did not reflect the finer-grained warning areas in AEMET's feed, so I had to rethink the subscription and routing model. This turned out to be a significant architectural change, costing me around 100 Lovable credits to correct. It was an expensive reminder that understanding the underlying data model before development is just as important when building with AI as it is with a traditional development team.
The move from Lovable and Supabase to Cloudflare was another distinct phase: moving from a working initial build to an application I could operate, test, deploy and back up myself. AI tools accelerated implementation, but they did not remove the need to understand the notification process or test its behaviour.
Results and reflections
The service had 32 active subscribers across six regions as of 4 November 2026, without any dedicated marketing. It's a modest start, but it demonstrates that the service is being used beyond my own household. The next challenge is understanding how consistently subscribers engage with the alerts and whether there's an opportunity to grow the service.
The biggest learning was how different building a functional prototype is from trusting an operational service. Authentication, external data quality, email delivery, alert timing, backups and edge cases all matter once somebody relies on the result. That experience now informs how I assess AI-assisted builds for client work.
Continue exploring
Interested in how I built it?
The full story covers the initial build, the problems I encountered and what I learnt.
Have an idea or a process you want working?
I take ideas from spec to a working product, and the same approach applies to automation and AI projects inside organisations.






