
Weekend Creative Challenge: Wanderwise
An AI agent that plans the whole road trip — real weather, a mapped route, meal stops, family-friendly hotels, kids' play areas, timing, and safety checks — from just a start and a destination.
Weekend Creative Challenge: WanderwiseAn AI agent that plans the whole road trip — real weather, a mapped route, meal stops, family-friendly hotels, kids' play areas, timing, and safety checks — from just a start and a destination.
At a glance
| |
|---|---|
What it is | Wanderwise — an AI road-trip concierge |
Creative output | A narrated, weather-aware day-by-day itinerary rendered as an interactive dashboard |
AI | Amazon Bedrock — Nova Lite (Converse API), run as a tool-use loop |
Compute / hosting | A single AWS Lambda behind a Function URL |
IaC | AWS SAM · least-privilege IAM scoped to Nova Lite |
Ground-truth data | Real weather + geocoding via Open-Meteo (no API key) |
Cost | No database, no API keys, no API Gateway — Free-Tier friendly |
Vision & what the app does
Most trip planners stop at the destination. Wanderwise plans the journey. You enter where you're starting and where you're going, pick a date and a trip style (family, couples, solo, friends, or slow travel), and an AI agent composes the entire day: a mapped route, real weather at both ends, breakfast / lunch / dinner stops spaced along the drive, family-friendly hotels, kids' play areas, an hour-by-hour timeline, a packing list, and a layered safety check covering road, natural/seasonal, and area-level risks.
The creative output is the itinerary itself — an opinionated, narrated travel plan generated fresh for your exact route and conditions, rendered as an interactive dashboard with a live map, weather cards, a vertical timeline, and rated recommendation cards. Run the same trip a month apart and you get a different, weather-aware plan, because the agent reasons over live conditions each time.
How you built it
I split the app into two honest layers so the useful parts are always trustworthy:
- A ground-truth layer that is fully deterministic: geocoding and real weather come from the free Open-Meteo API (no key required), and driving distance/time come from a haversine estimator. The map and weather never depend on the model.
- An AI agent layer: Amazon Bedrock's Nova Lite running a Converse tool-use loop. The Lambda hands the agent the already-resolved facts, and the agent can call the same
geocode/weather/routetools for extra lookups before emitting a strict JSON itinerary.
The biggest design decision was not letting the model invent weather or coordinates. My first version had Nova produce everything — charming plans, fictional temperatures. Moving weather, geocoding, and distance into real tools fixed accuracy while keeping the creative narrative.
The main challenge was getting reliable JSON from a chatty model. I solved it by pinning an exact schema in the system prompt, stripping code fences, extracting the outer
{...} , and adding one low-temperature corrective retry on parse failure. A second gotcha: Bedrock needs two gates — an IAM bedrock:InvokeModel permission and console Model access enabled for Nova Lite. The whole app is one Lambda behind a Function URL: GET serves the UI, POST runs the agent — no API Gateway, no database, no keys.AWS services used / architecture overview
- AWS Lambda (Python 3.12) behind a Lambda Function URL — serves the single-page UI and the planning API from one endpoint.
- Amazon Bedrock — Nova Lite (Converse API, inference profile
us.amazon.nova-lite-v1:0) — the trip-planning agent.
- AWS SAM for infrastructure-as-code, with an IAM role scoped least-privilege to invoke Nova Lite only.
- Free external data: Open-Meteo geocoding + forecast; map tiles via OpenStreetMap/Leaflet.

-
Flow: Browser → Lambda Function URL → deterministic geocode/weather/route + Bedrock Nova agent loop → a response combining ground-truth data (map + weather) with the AI plan → rendered dashboard.
Application Screenshots :



AWS Service Used Screenshots:




What I learned
- Agents are only as trustworthy as their tools. Grounding weather, location, and distance in real APIs turned a cute demo into something genuinely useful — and taught me where to draw the line between "let the model reason" and "give the model facts."
- Structured output is a design problem, not a prompt trick. A pinned schema plus defensive parsing and a corrective retry beat hoping for clean JSON.
- A Function URL + one Lambda is a superb weekend stack. No API Gateway, no database, and the whole app fits comfortably in Free Tier.
Link to app / repo
- Source (GitHub): https://github.com/parthi1914/wanderwise
Weather, geocoding, and route distance are real; restaurants, hotels, kids' areas, and safety notes are AI-generated suggestions — verify hours and book before travelling.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article