Plandrift

Your AI travel assistant and planner.

Comment

Inspiration

Trip planning today feels broken. Most tools jump straight to generating an itinerary, but that’s not how humans actually plan.

In real life, we first clarify constraints like budget, duration, season, flexibility, and priorities. Only after that do we check feasibility and then slowly refine a plan.

While experimenting with LLMs, I noticed that many AI travel planners fail not because of lack of data, but because they skip the thinking process. They optimize for output, not reasoning.

Plandrift was inspired by the idea that planning itself is a process worth modeling, not just the final answer.

What it does

Plandrift is an AI-powered trip planning engine that focuses on how a plan is created, not just the result.

Instead of generating a generic itinerary, Plandrift:

  • Asks clarifying questions about constraints and preferences
  • Breaks planning into structured phases
  • Generates a day-by-day plan that reflects assumptions and trade-offs
  • Treats planning as an iterative, evolving process

The goal is to move away from one-shot AI responses and toward something that feels closer to how humans actually plan trips.

How we built it

Plandrift was built as a full-stack application with a clear separation between planning logic and presentation.

Tech stack:

  • Frontend: Next.js + TypeScript
  • Backend: FastAPI
  • AI layer: Gemini models for reasoning and planning
  • Database: PostgreSQL for storing structured planning states

On the backend, the planning process is modeled as distinct phases (clarification, feasibility, assumptions, planning, refinement). Each phase produces structured outputs instead of raw conversation logs.

The frontend acts as a thin layer to guide users through the flow and visualize the generated plan.

Challenges we ran into

  • Modeling planning as a process instead of a single prompt was harder than expected
  • Deciding what to store as structured data versus transient reasoning required multiple iterations
  • Time constraints meant prioritizing core logic over UI polish
  • Handling incomplete or ambiguous user inputs without falling back to generic responses

Balancing speed (hackathon constraints) with architectural clarity was a constant trade-off.

Accomplishments that we're proud of

  • Designing a phase-based planning architecture instead of a one-shot generator
  • Shipping a working end-to-end prototype within a short timeframe
  • Building a system that asks better questions instead of pretending to know everything
  • Keeping the planning logic explicit and inspectable

Even in prototype form, Plandrift demonstrates a fundamentally different approach to AI planning.

What we learned

  • Good AI outputs depend more on problem framing than model size
  • Structured reasoning beats long unstructured prompts
  • Users trust systems more when assumptions are made explicit
  • Planning is an underrated but powerful application of agent-style workflows

Most importantly, we learned that slowing the AI down to think often produces better results.

What's next for Plandrift

  • Improve UI stability and clarity
  • Add iterative refinement where users can modify individual planning phases
  • Introduce versioning so plans can evolve over time
  • Explore real-time integrations for pricing and availability
  • Extend the planning engine beyond travel into other domains

Plandrift is still early, but the long-term vision is to rethink how humans and AI collaborate on planning.

Built With

Share this project:

Updates