---
type: "post"
id: "shipping-doordash-s-first-production-transformer-for-homepage-f97659"
slug: "shipping-doordash-s-first-production-transformer-for-homepage-f97659"
title: "Shipping DoorDash’s First Production Transformer for Homepage Store Ranking Deep Learning Engineering"
description: "DoorDash says it launched its first production transformer for homepage store ranking in mid-2025, adding sequence encoding to an established multi-task ranker rather than replacing it."
retrieval_nugget: "DoorDash says it launched its first production transformer for homepage store ranking in mid-2025, adding sequence encoding to an established multi-task ranker rather than replacing it. A three-week global A/B test improved order rate and merchant trial rate, with the strongest gains among lower-frequency and data-sparse customers; the absolute annualized order-value impact remains undisclosed."
published_at: "2026-11-12"
updated_at: "2026-11-12"
record_date: "2026-11-12"
date_kind: "published_at"
topics: ["models"]
entities: ["DoorDash","Store Ranker","Dipali Ranjan","Michael Chen","Yu Zhang","HSTU","MTGR"]
related_patterns: []
novelty_label: "new_fact"
what_changed: "DoorDash says its first production transformer for homepage store ranking launched in mid-2025, improving recommendation quality and engagement, especially among less-frequent customers."
prior_refs: []
source_url: "https://careersatdoordash.com/blog/homepage-store-ranking-transformer"
source_urls: ["https://careersatdoordash.com/blog/homepage-store-ranking-transformer"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:shipping-doordash-s-first-production-transformer-for-homepage-f97659"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/shipping-doordash-s-first-production-transformer-for-homepage-f97659/","markdown":"https://newruntime.com/posts/shipping-doordash-s-first-production-transformer-for-homepage-f97659.md","json":"https://newruntime.com/posts/shipping-doordash-s-first-production-transformer-for-homepage-f97659.json"}
markdown_url: "https://newruntime.com/posts/shipping-doordash-s-first-production-transformer-for-homepage-f97659.md"
---

# Shipping DoorDash’s First Production Transformer for Homepage Store Ranking Deep Learning Engineering

## Retrieval answer

DoorDash says it launched its first production transformer for homepage store ranking in mid-2025, adding sequence encoding to an established multi-task ranker rather than replacing it. A three-week global A/B test improved order rate and merchant trial rate, with the strongest gains among lower-frequency and data-sparse customers; the absolute annualized order-value impact remains undisclosed.

## Summary
DoorDash says it launched its first production transformer for homepage store ranking in mid-2025, adding sequence encoding to an established multi-task ranker rather than replacing it. A three-week global A/B test improved order rate and merchant trial rate, with the strongest gains among lower-frequency and data-sparse customers; the absolute annualized order-value impact remains undisclosed.

## What happened
DoorDash’s November 11, 2026 engineering announcement describes a homepage recommendation pipeline that retrieves eligible stores, scores them, blends store rankings within carousels, ranks carousel groups, and assembles the final page. The store ranker estimates how relevant each candidate is to a customer in the context of the current request. Its inputs include customer behavior, time and day, customer-store interactions, store quality and operations, and behavioral sequences. The ranker’s predictions are personalized signals for ordering stores within store-based carousels; other stages also shape the homepage.

The earlier ranker was a multi-task, multi-label deep neural network built around multi-gate mixture-of-experts (MMoE) and a deep-and-cross network (DCN). DoorDash says it combined behavior, customer-store affinity, store attributes, and real-time context to predict outcomes such as engagement and conversion. Histories mainly entered as aggregates—counts, recency buckets, and pooled embeddings—which could obscure the order and timing of actions. The transformer added a sequence-encoding branch to the existing system, combining customer-history representations with candidate-store and request features before the established task heads.

The company says the work began with full-journey attribution across impression, click, add-to-cart, checkout, and delivery, collected across the app rather than only the homepage. Click and conversion histories were restructured as sequences with temporal context. Click histories were capped at 500 interactions to limit feature-fetching latency; conversion histories were shorter and remained uncapped. For continuous features, DoorDash used percentile buckets—about 100 per feature—and computed percentiles over non-default values for features that were zero for most customers. It applied this conversion selectively to important features to balance accuracy and latency.

To fit the existing serving infrastructure, DoorDash changed the transformer block to use SwiGLU activations, RMSNorm, and pre-normalization. It reports a 0.08% relative AUC lift from those updates. Store IDs span millions of values, so frequency-based hashing and shared embedding tables reduced the model from about 160 million parameters to 10 million, a 94% reduction. The company reports relative offline AUC lift over the previous ranker rising from 0.88% to 1.15%. In a roughly three-week global A/B test, it saw improved overall order rate, especially among lower-frequency customers, and increased merchant trial rate in newer categories. DoorDash says these gains produced a material annualized order-value impact but withholds the absolute figure.

## What the archive adds
New Runtime’s earlier coverage described DoorDash separating orchestration from computation as a way to change the scaling boundary in a time-critical optimization system ([October 22, 2026 note](https://newruntime.com/posts/how-doordash-built-a-next-generation-real-time-eb6790/)). This ranking transformer is a new event from the same company, not an update to that dispatch system: it adds a separate production example of architecture shaped by serving constraints.

## Trend signal
This case supports the broader engineering lesson in New Runtime’s earlier DoorDash coverage: production architecture is shaped by workload and operational constraints, not model choice alone ([earlier note](https://newruntime.com/posts/how-doordash-built-a-next-generation-real-time-eb6790/)). It does not materially support one of the patterns New Runtime tracks; it is a concrete recommendation-system case rather than evidence of a broader agent or interface shift.

## Why now
DoorDash says the transformer launched in mid-2025 and the announcement now recounts the foundation, production constraints, and follow-up work. The post presents the launch as a step toward generative recommendation, while distinguishing that work in progress from the deployed pointwise ranker.

## Why it matters
For teams ranking large candidate sets, the case shows why adding a transformer is not simply a model swap. DoorDash first rebuilt attribution and sequence features, then placed a sequence encoder alongside an established multi-task system and retained candidate and request features. That lets a team test whether order, recency, and context improve decisions without discarding existing ranker behavior.

The results are encouraging but specific to DoorDash’s marketplace and system. The strongest reported gains were for users with sparse histories, where aggregate features provide less signal. The parameter reduction and explicit history cap also show that latency and model size were design constraints, not afterthoughts. Teams can reuse the evaluation questions—does sequence context improve online outcomes, for whom, and at what serving cost?—but should not assume the same lift transfers to their own data.

## Limits and uncertainty
DoorDash does not disclose the absolute annualized order-value impact, the experiment’s sample size, confidence intervals, or exact effect sizes for order and trial rates. The post reports relative offline AUC lifts and qualitative online wins, but does not provide enough detail to independently assess statistical significance or reproduce the experiment. Results come from DoorDash’s marketplace and serving environment; the company does not establish that the same architecture or gains will generalize to other ranking systems. Its generative-recommendation direction is described as active development with encouraging offline gains, not a production launch.

## Skill to build now
Audit whether your ranker’s aggregate history features hide meaningful order or timing. Select a slice with sparse user histories, build a sequence representation with timestamps or temporal buckets, and compare it against the current baseline in offline evaluation before testing online. Set a history cap and feature-fetch latency budget up front; DoorDash capped clicks at 500 interactions and converted only selected high-importance continuous features.

## Reader takeaway
DoorDash’s reported gains came from giving an established ranker access to ordered, time-aware behavior—not from replacing the whole recommendation stack. The practical lesson is to treat sequence modeling as a systems change: improve the journey data, fit the serving budget, and verify outcomes online, especially for users whose histories are sparse.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Models](https://newruntime.com/topics/models.md): Explore the models topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)