TasteRoute AI

Inspiration

Travel planning can become a long search through generic lists. TasteRoute AI gives people a quieter way to discover places that match their interests, with a small shortlist and room to choose their own pace.

What it does

Choose a city, interests, days and pace. Qloo resolves interests into typed tags; you explicitly choose their meaning. The app fetches real places from Qloo, preserves their original descriptions and addresses, and offers map links and a downloadable list. English and Polish interfaces are available.

A working agent, on your device

WebLLM runs Qwen2.5-0.5B-Instruct locally in the visitor's browser. The model chooses the recommend_places tool, the server obtains actual Qloo results, and the model then chooses select_places with IDs from those results. The visible trace shows the real decisions. Unknown IDs and duplicate selections are rejected. No model-service key or paid inference is needed.

How it is built

JavaScript, WebLLM, Qwen and a Worker-compatible backend deployed through Sites. Qloo's hackathon API is central: /v2/tags and /v2/insights supply the real discovery data. The Qloo key stays in a private server secret. Tag selections are signed and expire; requests have validation, response-size limits, timeouts and per-isolate throttling. GPU work is serialized to prevent overlapping generations. AI coding assistance was used to implement and debug the project, with participant feedback on the experience.

Challenges and lessons

Taste signals are not hard category filters. Coffee uses an explicit tag filter, while other interests remain affinity signals. We also pinned the official model revision to make downloads reproducible and fixed hosted request compatibility. Keeping the model's choices bounded by provider IDs avoids presenting invented places as recommendations.

What was checked

Thirteen automated checks cover the agent, model queue and API with injected responses. Separately, a genuine public Chrome session completed Qwen -> Qloo -> Qwen for Lisbon / Coffee, selecting Matoli Gelato from live API results. This is one verified scenario, not a claim that every destination or provider failure has been tested live.

Try it

Open the public demo in current desktop Chrome or Edge with WebGPU and approximately 1 GB of available GPU memory. Choose Lisbon, Coffee, then Find my interests and Coffee / urn:tag:cuisine:qloo. Start free AI, wait for AI is ready on this device, and choose Let AI choose my places. First use downloads about 290 MB of model files plus runtime assets and may take several minutes. Show places from Qloo works independently of local AI.

Current boundaries and next steps

The prototype offers four cities and visit ideas rather than a timed route. Check opening hours, availability, prices and distances independently. Qloo affinities describe aggregate tastes, not certainty about an individual. The original illustrative demo is retained separately as sample data. Future work includes more destinations and verified practical travel information.

Public demo: https://tasteroute-ai.stillmole8.chatgpt.site MIT source and setup: https://github.com/jacman1988-super/tasteroute-ai

Built With

• javascript
• qloo
• webllm