Ski-Tripper

Ski Tripper helps a ski group find and agree on a great ski destination without the “entertaining” chaos of a WhatsApp debate.

Built-in AI uses everyone’s ski holiday preferences to guide searches of an enriched catalogue and generate narrative assessments on who will (or won’t) enjoy a resort; everyone creates proposals and enters them into voting rounds to reach a collective decision.

At a technical level, I created it to gain hands-on experience with LLMs. Although a small application, it exercises the full stack and lifecycle so gives me ideas about how AI should be used on serious, large-scale software projects. In particular, it covers:

• The frontend (essentially React).
• The backend (small TypeScript server and pocketbase).
• Full set of unit tests and end-to-end tests (using Playwright).
• Fully automated server build, configuration and deployment to DigitalOcean (using xec).

LLMs were used throughout the project:

1. To build the application. I experimented with a number of agentic tools, models, local or cloud-based providers, and configurations (skills, MCPs, etc), settling on OpenCode and open source models running in Ollama cloud so I can track new models and updates. Much of ski-tripper was written with the help of GLM5.1.
2. To create a rich catalogue of resorts with standardised fields and descriptions. This involves a pipeline which seeds a list, enriches from qualified sources, assesses quality, and fixes inconsistencies using an independent model.
3. To make it easier for users to search the catalogue of resorts. An embedding model is used to one-time create embeddings for each resort as part of catalogue generation, and then the same model is used again in the client browser to quickly find similar resorts.
4. To generate resort search text from participant preferences. An LLM is fed everyone’s preferences and instructed to generate search text to run against the embedding model (previous), and so make it easier to find candidate resorts the group will enjoy.
5. To assess a proposal against the likes/dislikes of the participants. An LLM is used to create a narative assessment of the match between a proposal and the likes/dislikes of the participants, trying to identify who would especially like a resort and who might find it less appealing.
6. To automate the testing of the applicaton UI by performing user interactions using a headless browser, looking for bugs and increasing confidence that the application behaves in a reasonable way.

For technical details, see the project’s GitHub repository.