TripFork AI Trip Planner
Tool-augmented LLM planner that decomposes natural language travel goals into multi-step itineraries, grounds each recommendation with live API data, and resolves constraint conflicts across days and budgets.
Skills involved
What This Is
A tool-augmented LLM planner for travel. The core challenge isn't generating itinerary text any LLM can do that. The challenge is grounding the plan in real-world constraints: actual flight times, real restaurant hours, genuine transit durations, accurate weather forecasts.
Ungrounded travel plans look coherent and are useless. A plan that suggests a 2-hour museum visit followed by a 1-hour transit in 45 minutes is worse than no plan at all.
The Planning Architecture
Trip planning decomposes naturally into sub-problems that current LLMs can handle well individually but struggle to coordinate:
- Preference extraction: parse the user's natural language goals into structured constraints (budget, travel style, must-see vs. nice-to-have, mobility requirements)
- Temporal constraint resolution: lay out the skeleton schedule flights, check-ins, fixed commitments and compute the actual free time available
- Activity selection: given free time slots and preferences, select activities with real operating hours, durations, and booking requirements
- Route optimisation: cluster activities by geography to minimise transit waste
- Conflict resolution: when constraints are incompatible, surface the trade-offs and ask the user to choose rather than silently ignoring one
Each step is a separate LLM call with tool access (maps API for transit times, weather API for forecast, search for opening hours and prices).
The Gemini Integration
Gemini's long context window makes it well-suited for the final synthesis pass: given the full constraint graph and resolved activities, generate a coherent narrative itinerary that reads naturally and includes actionable details (addresses, booking links, tips).
What We're Learning
- How to design tool-use protocols that handle tool failures gracefully (API down, no results)
- Whether explicit constraint decomposition improves plan quality vs. end-to-end generation
- How to represent "I'm flexible on X but not Y" in a way the planner can reason about
What's Next
- Real-time API integration (flights, hotels, activities)
- Multi-destination trip planning with connecting journeys
- Collaborative planning: multiple users, shared constraints, consensus itinerary
Last updated Oct 14, 2024
← Back to all projects