Logistics ops · Production

DoorDash 2025 Summer Intern Projects: GenAI Voice Agent, Storm Mode ETA, Probabilistic ETA Model, and LLM Alcohol Recommendations

The problem

DoorDash's ETA model was thrown off during high-stress events while live operators relied on unvalidated, untracked CSV files; the robocall system could only capture yes/no responses on the day of a closure rather than capturing detailed hours in advance; the Weibull-based ETA model's assumptions may not reflect true delivery time distributions; and the Alcohol Landing Page lacked personalized item carousels, limiting product discovery.

First attempt

The existing bulk CSV tool offered no historical evidence and forced the ETA team to query multiple databases for change history; the DTMF robocall system was limited to simple button-press responses and could not capture detailed hours in advance.

Workflow diagram · grounded in source
1
Voice call provider selection
trigger
“Scheduler picks the provider via factory”
2
AI voice agent merchant conversation
ai_action
“It guides the merchant to state whether they're closed, keeping regular hours, or instituting special hours, and then elicits exact times”
3
VAPI structured response to webhook
integration
“VAPI sends structured response to webhook”
4
Webhook triggers special hours update
output
“Webhook interprets and triggers special hours update in MenuDataService”
5
Kafka downstream notification
integration
“Kafka events notify systems downstream and keep stores open or closed”
6
Storm mode ETA auto-calculation
ai_action
“the system automatically calculates an area's ETA pad in real-time, calculating the time based on the specific area's supply-demand ratio and distance between the restaurant and consumer”
7
Store ETA Pad Auditor evaluation
feedback_loop
“uses empirical delivery-time distributions to estimate each store's ETA accuracy both with and without the current pad. The auditor also tests a range of candidate pad values to find the optimal OTA pad and classifies the current one as …”
8
Probabilistic ETA model prediction
ai_action
“Delivery durations are divided into bins — for example, 60-second intervals — and the model predicts the probability of each bin. Together, these form a probability distribution of arrival times”
9
LLM alcohol recommendation generation
ai_action
“taking in a user's order history and search terms, using semantic search to surface real, orderable products from the DoorDash catalog”
10
LLM carousel title judge and re-rank
ai_action
“we added a second step where the LLM evaluates its own candidates. It filters out overly editorial or generic titles and re-ranks the rest”
Reported outcome

DoorDash shipped a structured Logistics Console with storm mode automation and an ETA pad auditor, a GenAI voice agent that captures natural language merchant responses to update store hours, two probabilistic ETA model candidates that achieved statistically significant improvement in on-time arrival in an A/B test, and an LLM-powered alcohol recommendation carousel with generated titles.

Reported metrics
ETA system safety and transparencysignificantly improved the safety, transparency, and agility of DoorDash ETAs
On-call response speedspeeds up on-call response and debugging
Storm mode operator manual effortremoves manual guesswork for live operators
on-time arrival (OTA) in A/B teststatistically significant improvement in OTA and earliness
Show all 5 reported metrics
ETA system safety and transparencysignificantly improved the safety, transparency, and agility of DoorDash ETAs
on-call response speedspeeds up on-call response and debugging
storm mode operator manual effortremoves manual guesswork for live operators
on-time arrival (OTA) in A/B teststatistically significant improvement in OTA and earliness
ETA model MAE and CRPSimproving both MAE and CRPS
Reported stack
VapiKafkaLarge Language ModelsEntityCache
Source
https://careersatdoordash.com/blog/part-4-doordash-2025-summer-intern-projects/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

DoorDash shipped a structured Logistics Console with storm mode automation and an ETA pad auditor, a GenAI voice agent that captures natural language merchant responses to update store hours, two probabilistic ETA mod…

What tools did this team use?

Vapi, Kafka, Large Language Models, EntityCache.

What results were reported?

ETA system safety and transparency: significantly improved the safety, transparency, and agility of DoorDash ETAs; On-call response speed: speeds up on-call response and debugging; Storm mode operator manual effort: removes manual guesswork for live operators; on-time arrival (OTA) in A/B test: statistically significant improvement in OTA and earliness (source-reported, not independently verified).

What failed first in this deployment?

The existing bulk CSV tool offered no historical evidence and forced the ETA team to query multiple databases for change history; the DTMF robocall system was limited to simple button-press responses and could not cap…

How is this logistics ops AI workflow structured?

Voice call provider selection → AI voice agent merchant conversation → VAPI structured response to webhook → Webhook triggers special hours update → Kafka downstream notification → Storm mode ETA auto-calculation → Store ETA Pad Auditor evaluation → Probabilistic ETA model prediction → LLM alcohol recommendation generation → LLM carousel title judge and re-rank.

Related logistics ops cases