FA
Separate three decisions
Demand forecasting estimates where and when orders may arise. Routing chooses feasible sequences subject to vehicle, time and access constraints. Warehouse slotting assigns storage positions. These are related optimization problems, but one model does not necessarily solve all three.
Treat this architecture as illustrative
A proposed workflow may combine historical orders, permitted location data, weather and live operational updates. Check coverage, freshness and missing data before relying on a feed. This article does not establish that Mercado Libre or Rappi deploys this exact architecture, uses a particular learning method or operates it in every market.
Measure the trade-offs
Compare forecasts with a simple seasonal baseline and routes with the existing dispatch process. Track delivery reliability, distance, cost, failed deliveries, worker burden and customer outcomes. A dynamic route can create disruption as well as savings. Evaluate changes in a bounded pilot before expanding.
Do not assume transferability
A model fitted to one city can fail in another because addresses, demand, access rules and operating practices differ. Local data may be valuable, but neither accumulated data nor infrastructure complexity proves global superiority. Revalidate performance and permissions in every new deployment context.
Sources and scope
Practitioner analysis by Fernando Angulo. Proposed examples are not measured outcomes.