Skip to content
Roamly

How the AI works

Built on language models. Built to switch between them.

Roamly is model-agnostic by design. It routes planning, adjusting, and triage across the right model for each task, through one provider layer, with fallback if a provider is unavailable.

The pipeline

Three stages, one route

Every Roamly trip request goes through three stages. Each stage is routed to the right model for that kind of work, and the routing is overridable without a code change.
01

Understand your style and constraints

Roamly reads your start, your rough destination or a loop, how many days you have, and how you travel, including kids, a dog, RV, or tent. It identifies what matters most for pacing and what cannot be moved. A lightweight model classifies the request before it is routed to the right engine for the actual planning.

02

Build a paced route

The language model reasons about the whole trip at once, not day by day. It weighs how far to drive each day for your travel style, where there is somewhere good to sleep, which detours are worth the extra time, and what to flag as confirm-before-you-go. The output is a full day-by-day route, not a list of links to open in twenty tabs.

03

Adapt as you tweak

When you change a day, ask for a shorter drive, or add a stop, Roamly regenerates only the days that need to change and re-flows the rest. It holds the parts you liked and adjusts only what moved, so the whole trip still makes sense after a tweak.

Model routing

What runs where, and why

This table is generated from the live routing configuration. Every field is real: the model named is the one handling that task in this deployment. Each route is overridable by environment variable, so a deployment can move a step without a code change.
TaskModelProviderTierApprox cost / M tokens

Plan a full day-by-day route

plan

gpt-4.1-miniopenaibalanced$1.60

Adjust the days affected by a tweak

adjust

gpt-4.1-miniopenaibalanced$1.60

Quick routing triage

classify

gpt-4.1-nanoopenaifast$0.40

Roamly is model-agnostic and routes across leading models through one provider layer, with a fallback chain so one lab being unavailable never loses a request. OpenAI is wired live in this deployment. Anthropic, Google, Meta, and Mistral adapters are declared and switch on when their respective API key is present. Long trips with many days escalate automatically to the frontier model.

Declared candidates

gpt-4.1

frontier

openai

$8.00 / M tokens

long trips with many days

claude-sonnet

frontier

anthropic

$9.00 / M tokens

careful pacing and honest tradeoffs

gemini-flash

fast

google

$0.50 / M tokens

fast first drafts

llama-open

open

meta

$0.30 / M tokens

self-hosted planning for privacy-sensitive deployments

Honesty and limits

What Roamly says plainly

Roamly is useful because it is direct about what it does not know, what to confirm, and what it is doing with your data.

Suggestions, not certainties

Roamly's routes and stops are suggestions. Hours, road conditions, and campground availability change. Closures happen. Every route includes plain notes on what to confirm before you rely on it, and Roamly never inflates its confidence to seem more useful than it is.

Recommendations stay honest rather than sponsored-feeling. A detour appears because the judgment is that it is worth the extra time, not because a placement fee was paid.

Your trip data stays yours

When Roamly builds a route, your trip details are sent to a model provider only to produce the response. They are not used to train models, stored beyond the request by the provider, or sold. The provider agreements require this in writing.

Your trips live in Roamly's own database, not the provider's. You can export any route or delete any trip and your whole account from Settings, which removes it from the database. The Your data and Security pages have the specifics.

Ready to see the route?