Dynamic Moments: Weaving LLMs into Deep Personalization at DoorDash

The landscape of e-commerce personalization is undergoing a significant transformation, with DoorDash at the forefront of this evolution. The food delivery and local commerce giant is pioneering a hybrid approach that combines large language models (LLMs) with traditional machine learning to create dynamic, moment-aware personalization experiences. This shift represents a fundamental change from static merchandising strategies to a more nuanced understanding of consumer behavior that adapts to short-lived user intent and massive catalog abundance.

The Evolution of Personalization

Personalization in e-commerce has evolved significantly over the past two decades, starting with matrix factorization approaches similar to the Netflix Prize, moving through LDA-based methods, and eventually embracing deep learning techniques like wide plus deep models, MTML, and two-tower embeddings for retrieval. While these "classic" personalization systems effectively learned from engagement and product metadata, they struggled with several limitations:

1. Static understanding: They couldn't react quickly enough to changing user interests
2. Limited context: They couldn't incorporate real-time moment awareness
3. Narrow knowledge: They lacked the world knowledge embedded in modern LLMs
4. Limited explainability: They couldn't articulate recommendations in natural language

"The problem with classic personalization is that it doesn't always meet the customer at the moment of need," explains Sudeep Das, Head of Machine Learning and Artificial Intelligence at DoorDash. "If it's late night and you're hungry, wanting snacks, understanding that in the moment context and making personalization be dynamic in that context is something that classic personalization cannot do very well."

The Architecture of Dynamic Personalization

DoorDash's approach to hyper-personalization is built on a sophisticated architecture that combines several key components:

Product Understanding Layer

DoorDash has made a significant shift from human-centric systems to AI-driven product understanding. This transformation has dramatically improved efficiency:

"There's a task that we used to call extraction," Das notes. "For a headphone, we needed to extract whether it's noise canceling, what's the brand, what's the color, is it over-the-ear or an earbud. This task used to take 28 days; we now do it in 2 days today."

This acceleration was achieved through:

• Fine-tuned LLMs for specific extraction tasks
• Grounding systems with RAG (Retrieval-Augmented Generation) using DoorDash's ontology and taxonomy
• Agentic processes to handle highly abbreviated data from small merchants

Consumer Profiles through LLM Narratives

A core innovation in DoorDash's approach is the use of LLMs to generate natural-language consumer profiles. These profiles replace traditional feature vectors and embeddings with expressive, narrative descriptions of user preferences and behaviors.

"We represent consumer profiles in plain English," explains Pradeep Muthukrishnan, Head of Growth for New Business Verticals at DoorDash. "You can write anything you want about what you think this user is interested in, their preferences when it comes to stores, categories, item level, brand level preferences. This has enough expressiveness to capture nuanced consumer behavior."

These profiles are organized into memory blocks covering:

• Dietary habits
• Household information
• Category preferences
• Item and brand preferences
• Taxonomy preferences

The profiles serve as shared primitives used across different applications, from notifications to homepage carousels to search ranking.

Content Blueprint Generation

Using these consumer profiles, DoorDash employs LLMs to generate "content blueprints" for personalized merchandising. These blueprints outline what carousels and collections should be shown to each user for different use cases:

• Evergreen use cases (grocery stock-up, brunch basics, nightly snacks)
• Moment-specific use cases (Black Friday, back to school, flu season)

The system outputs both the carousel structure and the specific queries needed to populate them with relevant items. Additional constraints can be added based on price sensitivity, brand preferences, or merchant affinities.

Real-Time Blending with Traditional ML

While LLMs excel at generating profiles and content blueprints, DoorDash wisely retains traditional machine learning for certain tasks:

"We don't want LLMs in your serving path itself due to latency and cost concerns," Muthukrishnan notes. "Most content generation happens offline, but real-time population of carousels needs to respect inventory, current deals, and any changes in user intent since they opened the app."

The "DoorDash Brain" service orchestrates this blending by:

• Streaming consumer events
• Maintaining and serving consumer profiles
• Tracking evolving user intent through embeddings
• Combining offline-generated content with real-time signals

The Ideal Experience: Alice's Black Friday Shopping

To illustrate the impact of this system, consider Alice, a DoorDash user with a history of ordering electronic essentials and showing interest in high-end headphones. During Black Friday:

1. Alice opens the DoorDash app to find a homepage flooded with electronic deals
2. The system has identified her interest in noise-canceling, over-ear headphones based on recent browsing behavior
3. The page includes a diverse mix of relevant items (not just headphones) to encourage exploration
4. Items are shown from merchants Alice has affinity toward (Best Buy, Home Depot)
5. The content dynamically adapts if Alice's intent shifts during her session

Evaluation and Optimization Challenges

Evaluating this highly personalized system presents unique challenges. Unlike traditional cohort-level merchandising where humans could evaluate 30 different use cases and 100 carousels, DoorDash now generates 50 unique carousels per user with different copy and items.

DoorDash evaluates the system across three axes:

1. Quantitative metrics (click-through rates, conversion rates)
2. LLM-as-a-judge assessments
3. Human annotated feedback

They use GEPA (Genetic-Pareto optimization) within DSPy to optimize the compound AI system. This evolutionary approach treats prompts, profiles, retrieval logic, and ranking objectives as parameters to be optimized based on the reward function derived from evaluation metrics.

Lessons and Insights

Through implementing this system, DoorDash has learned several valuable lessons:

The Power of Hybrid Approaches

"LLMs shine at turning messy behavior into clean, understandable narratives," Das observes. "Deep learning models excel when optimizing for concrete metrics like CTR or CVR under constraints. Don't let your LLMs do the last-mile ranking—let them handle content ideation while your ML models handle ranking."

Organizational Considerations

Several organizational factors contributed to the success of this initiative:

• Investment in shared primitives (profiles, product graphs, evaluation frameworks)
• Treating LLM + deep learning integration as product work, not just infrastructure
• Starting small with concrete experiments before building generic frameworks
• Focusing on measurable returns on investment

The Future of Agent-Based Experiences

DoorDash is actively developing more agent-based experiences that leverage these personalization primitives:

"We're developing agents that will help the Dasher shop as their insider store," Das reveals. "The same profile primitive we built can be injected into the agent's context through RAG, enabling last-mile reasoning over which items to pick based on user preferences and real-time constraints."

User Impact and Industry Implications

For users, this approach means shopping experiences that feel increasingly intuitive and anticipatory. The system moves beyond "people like you like X" to "you need X now"—recognizing and responding to immediate context and intent.

For the industry, DoorDash's implementation demonstrates a pragmatic approach to LLM integration that balances the strengths of generative AI with traditional machine learning. It shows how companies can leverage LLMs not just for content generation, but for creating structured, actionable intelligence about users and products.

As e-commerce continues to evolve, this hybrid approach may become the standard for personalization systems that need to balance real-time responsiveness with deep understanding of user preferences across massive catalogs.

For more information about DoorDash's approach to personalization, you can explore their engineering blog or learn more about their machine learning initiatives.