Generative AI app · Case study
AI Travel Concierge
Enter a destination, travelers, duration, budget, focus and must-do list, and get a full day-by-day itinerary. Curated local knowledge grounds the plan, so the model works from real tips instead of guessing.
- Role
- Designer and builder
- Stack
- Python, Streamlit, Gemini, Groq / Llama fallback, Docker
- Data
- Synthetic, for demonstration
- Status
- Working prototype
Walkthrough coming soon
Loom recording in progress
What it does
A sidebar form collects the trip details. The app builds a planning package from agent instructions plus the traveler's requirements and sends it to Gemini to produce the itinerary. Results are also saved to disk.
Design choices
- Lightweight retrieval (RAG). Each supported city has a folder of curated highlights covering food, sights, neighborhoods and practical tips. When the typed destination matches a folder by exact, alias or fuzzy match, those highlights are injected into the prompt as reference material.
- Easy to extend. Adding a new city means adding a folder with a highlights file. No code changes.
- Reliability. If Gemini errors out, the app falls back to a Groq / Llama model automatically.
- Deployable anywhere. It runs locally, in Docker, or on Streamlit Community Cloud with keys held in secrets.
Why it is in my portfolio
It is the smallest of the three projects, and the clearest example of grounding a model in a curated knowledge source with a fallback path. That is the same pattern I apply in the finance agents.