Inspiration
Weather apps can tell travelers that rain, wind, heat, or thunderstorms are coming. They rarely answer the more useful question: what should change in a real trip without breaking the commitments that cannot move?
WeatherWise Trip Guardian turns a natural-language itinerary into one practical, validated repair. It is designed for travelers who need a clear decision—not a list of loosely related suggestions.
What it does
A traveler describes the trip in ordinary language, including fixed commitments, flexible activities, weather sensitivity, and whether city changes are allowed.
Guardian then:
- Uses Gemini on Vertex AI to interpret the itinerary into a typed trip intent.
- Fetches forecast evidence from Open-Meteo.
- Computes weather risk with deterministic Python rules.
- Compares same-day alternatives and rejects cities whose risk is equal to or worse than the original destination.
- Uses Gemini to propose a repair, then validates every change deterministically.
- Returns exactly one Recommended repair, prioritizing safety first and minimum disruption second.
- Lets the traveler accept the repair and build a canonical, date-grouped final itinerary.
- Generates an in-memory PDF and RFC 5545 calendar download from that accepted itinerary.
The current product follows a simple mental model: one destination city per day. On a dangerous flexible day, Guardian can replace that whole day with the best genuinely safer city. Fixed commitments remain fixed.
How we built it
The public experience is a Streamlit application deployed on Google Cloud Run. The runtime combines:
- Gemini through Vertex AI for semantic intake and repair synthesis
- Pydantic schemas for typed model and tool boundaries
- Open-Meteo for forecast and geocoding data
- deterministic Python risk scoring and feasibility validation
- bounded Gemini correction followed by deterministic re-validation
- in-memory PDF and ICS generation
- a dedicated Cloud Run service account and Secret Manager-backed state signing
Gemini is an essential part of understanding and proposing a useful repair, but it never owns the safety verdict. Risk levels, fixed-activity preservation, alternative-city integrity, and final feasibility are enforced outside the model.
How Codex and GPT-5.6 were used
Codex with GPT-5.6 was the main engineering collaborator for the Build Week implementation. It helped transform an isolated inherited baseline into the current Guardian product by:
- designing the typed Gemini-to-validator pipeline;
- implementing and reviewing deterministic risk, alternative ranking, and repair validation;
- hardening untrusted prompt and provider boundaries;
- reducing the public flow from several competing strategies to one actionable recommendation;
- building deterministic final-itinerary, PDF, and calendar exports;
- creating regression, schema, prompt-injection, and integrity tests;
- analyzing performance and evaluation traces;
- preparing and verifying the Cloud Run release; and
- documenting the provenance boundary and major technical decisions.
The repository preserves the inherited baseline as an immutable tag and keeps Build Week work on a separate branch so judges can inspect the delta.
Challenges we ran into
The hardest problem was separating helpful model judgment from facts that must never drift. A fluent repair can still be unsafe if it moves a fixed event, recommends a worse city, invents missing forecast data, or quietly changes an activity without evidence.
We solved this with a strict sequence:
typed Gemini proposal → deterministic validation → bounded correction → deterministic re-validation
Another challenge was making the result understandable. Earlier versions exposed too many strategies and too much evidence. The final UI gives one recommendation, a compact before/after explanation, and an optional bounded evidence view.
Accomplishments that we're proud of
- A working public application with a complete intake-to-download flow
- Deterministic protection against risk downgrades and equal/worse alternative cities
- Fixed commitments preserved through recommendation, final itinerary, PDF, and ICS
- A ten-case evaluation in which all four deterministic integrity metrics reached 1.0
- Batched, cached weather retrieval that substantially reduces external requests
- A production deployment that remains separate from the inherited service
What we learned
The best use of an LLM in a safety-relevant planning product is not to let it own every decision. Gemini is strongest at understanding messy human intent and synthesizing a useful repair. Deterministic code is strongest at enforcing invariants and making the final answer auditable. Combining the two produces a result that is both more useful and more trustworthy.
What's next
Next steps include richer regional candidate coverage, official alert integrations, durable cross-instance session storage, and broader timezone validation. We also want to keep improving the explanation layer without exposing internal traces or overwhelming the traveler.
Try it
The live demo requires no account. Enter a near-term itinerary, allow city changes, and compare the original risk calendar with the single validated recommendation.