Problem
DoorDash restaurants start cooking when they confirm an order, but a courier may arrive much later or much earlier. Early cooking leaves food cooling on the counter; late cooking makes couriers (Dashers) wait. For quick-service restaurants the Auto Order Release (AOR) flow holds a confirmed order until the courier enters a geofence around the store. The fixed per-store release distance fit real prep and arrival times poorly, merchants complained about waiting couriers, and every new store needed a hand-drawn geofence.
Example from the article: larger orders at certain times take longer to prepare, yet a store on AOR held them until the courier was close, so the courier waited.
Solution in brief
The team moved the release decision from a per-store setting to a per-order one. First came a heuristic version, tested in production experiments. In parallel they built an ML model of food preparation time and used it to replace the heuristics, then to set the release distance.
Questions this case answers
- How to replace a hard-coded per-entity rule with a per-case ML decision without losing the production gains already delivered by a simple version
- How to tune a prediction model's loss when offline error does not translate into the business metric and a second metric must not degrade
- How to decide when to release an order to the kitchen so that couriers do not wait for food and the food does not go cold?
- How to ship an ML decision feature in stages, starting from a heuristic that already earns production results?
Tools and methods
- Rule-based release heuristics — per-order choice between instant release and delayed (AOR) release, for example by order size and time of day
- LightGBM prep-time model — predicts how long the restaurant needs to cook, compared with the courier's arrival estimate
- Production experiments — check the wait-time effect against food-quality and delivery-time guardrails
Google or a one-time email code. No password.