Patagonia Solar

From "could I go solar?" to a quote from an installer

Going solar or adding wind in Patagonia means solving three problems at once: working out what size of system you need, finding the equipment in the first place, and understanding the net billing and grid connection rules where you live. Patagonia Solar tackles all three in one place. Its hybrid calculator takes either your monthly consumption or a list of your appliances, uses local sun and wind data to recommend capacity, and then lets you request quotes from verified installers.

Behind the catalogue sits an automated pipeline that runs every month. It collects products, uses an LLM to extract their specifications, and merges those with datasheet values and values matched by pattern in order of trust, keeping only specs that make sense for that type of product. Results sync to Postgres and to a vector index, which the AI assistant searches through Gemini tool calls to find equipment, installers and suppliers by country and region. The pipeline also tracks its own time budget, so it stops cleanly before the CI job's time limit.

The assistant has a page of its own, behind sign in, and it is deliberately not free to improvise. Every renewable energy question has to go through a catalogue search tool, whose arguments are checked against a schema before it runs and whose results are read back out of the database, so an answer describes equipment, suppliers and installers the site actually holds rather than something the model remembers. If it tries to reply without searching, the server runs the search anyway; if a search comes back empty, it offers better questions instead of guessing; and results older than a year are left out. That turns the chat into a conversation someone can actually plan around, in Spanish or English, at whatever point of the switch to renewable energy they are at.

Several pieces are built ahead of launch, including an installer directory, supplier quotes and group buying pools, where buyers combine their orders, suppliers bid with sealed quotes, and everyone is refunded if a pool misses its minimum number of participants by the deadline. Community reviews, machine curation scores and supplier specs are kept as three separate signals so none can distort the others, and structured data for search engines is validated in CI on every push.

• 850+Products in the catalogue across five equipment categories
• 51Data models behind the platform
• 2Countries covered, with sun and wind data by province or comuna