AI product
Route-Master
A route planning assistant built on the Anthropic Claude SDK that takes a day's stops and turns them into a sensible route, on a live map, behind a subscription.
The problem
Solo business owners (the people running a delivery van, a catering round or a repair service on their own) plan their day on paper or in their head. Enterprise route optimisation software exists, but it is priced and shaped for fleets, not for one person with a phone and a list of stops.
Route-Master is built for that person: put the day's stops in, get a route back that makes sense, on a map.
What I built
- An AI planning layer on the Anthropic Claude SDK that reasons about the day's stops and returns an ordered, practical route rather than only a raw distance calculation.
- An interactive Leaflet map so the plan is something you look at and adjust, not a list of addresses to interpret.
- Stripe subscription billing for the paid tier, wired the same way as the rest of the product rather than bolted on at the end.
- A Supabase backend for auth and storage, with Zustand handling client state and Framer Motion carrying the interface transitions.
- Zod schemas validating both user input and the model's structured output, so a bad response never reaches the map.
Architecture decisions
Treat the model's output as untrusted input
Everything coming back from the LLM is parsed and validated against a Zod schema before the application uses it. A language model is a very capable component, and like any external service it needs a contract at the boundary.
Keep the map authoritative for what the user sees
The AI proposes, the map presents, and the user decides. Route planning is a domain where local knowledge beats any model, so the interface is built for adjusting the plan, not just accepting it.
Next.js 16, with types all the way through
Built on Next.js 16 and TypeScript end to end, from the map components down to the API handlers, so a shape change surfaces at build time rather than in production.
Hard parts
- Making AI output reliable enough to build a product on. Prompting is only half of it: the schema validation, the retries and the fallbacks are what make the feature dependable.
- Latency versus quality. A planning call that thinks for too long feels broken, so the interface has to keep the user informed while the work happens.
- Map interaction on a phone. The people this is built for are standing next to a van, not sitting at a desk.
Outcome
Route-Master is a working, deployed product that solves a real operational pain point for a specific kind of business, and it is the project where I learned to build with an LLM as a dependable component of a system rather than as a novelty.