−30% in travel cost across 70+ field routes

70+ routes across Brazil, walked by people who visit stores to audit them. The routes had to be optimised without ignoring what field work really looks like.

Modelling the problem with the operations team, choosing the solver and connecting it to the real cost of travel.

The problem is a VRPTW, so I modelled it with Google OR-Tools (this was before LLMs), adding penalties, visit priority by coordinate and real capacity limits: working hours, lunch, first and last store, the trip home. Public transport cost came from Google APIs per city, and where there was no data, from the average fare by postal code — so a route could be checked for economic sense, not only geographic sense.

A classic solver means modelling every constraint by hand, which is slower to build. I took that on because the flexibility to tune each parameter mattered more than getting there fast.

−30% in real travel cost compared to what was being spent before.