Logistics ops · Production

DoorDash uses ML and mixed-integer optimization to solve last-mile dispatch with DeepRed

The problem

DoorDash needed to solve a complex three-sided dispatch problem — matching each order to the right Dasher at the right time while accounting for geography, timing, batching opportunities, supply/demand imbalance, weather, and traffic.

Workflow diagram · grounded in source
1
New order arrives
trigger
“When a new order arrives to our dispatch engine, we first update our understanding of the current state of the marketplace and how this order interacts with the Dashers and other orders.”
2
Offer candidate generation
ai_action
“By looking at the available Dashers and other orders, we are able to construct potential offers for our new order: a set of Dashers that this order could be offered to and possibly other orders that could be picked up by the same Dasher.”
3
ML layer predictions
ai_action
“With a set of potential offers in hand, we are ready to make some estimates using our ML models, including, but not limited to: order ready times, travel times, and offer acceptance likelihood.”
4
Optimization layer dispatch decision
ai_action
“The optimization model does the work of scoring and ranking potential offers, making decisions about batching orders, and strategically delaying dispatches when necessary.”
5
Order offered to Dasher
output
“The order is offered to the Dasher we have chosen, and we wait to see if they will accept or decline the offer. If necessary, we will find another Dasher to offer the order to, until the order is picked up at the store and delivered to t…”
6
Continuous model improvement
feedback_loop
“we continuously retrain our ML models and use rolling historical and real-time features that make sure the inputs to our models stay fresh”
Reported outcome

DoorDash built DeepRed, an ML-plus-optimization dispatch engine that handles millions of daily orders, continuously improves through simulation and experimentation, and balances efficiency for Dashers with on-time delivery for consumers and merchants.

Reported metrics
Continuous dispatch improvement rate1% better every day
Reported stack
DeepRedGurobiSibyl ML platform
Source
https://careersatdoordash.com/blog/using-ml-and-optimization-to-solve-doordashs-dispatch-problem/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

DoorDash built DeepRed, an ML-plus-optimization dispatch engine that handles millions of daily orders, continuously improves through simulation and experimentation, and balances efficiency for Dashers with on-time del…

What tools did this team use?

DeepRed, Gurobi, Sibyl ML platform.

What results were reported?

Continuous dispatch improvement rate: 1% better every day (source-reported, not independently verified).

How is this logistics ops AI workflow structured?

New order arrives → Offer candidate generation → ML layer predictions → Optimization layer dispatch decision → Order offered to Dasher → Continuous model improvement.

Related logistics ops cases