Presentation: From Models to Agents: Building Context-Aware Consumer AI at Scale at DoorDash

The shift DoorDash is undertaking, as detailed by Sudeep Das, represents a crucial evolution in how businesses leverage AI for personalized experiences. Moving beyond static, one-off predictions to an agentic recommendation platform is a significant architectural change, and one that many companies will need to consider. The core of this transition lies in embracing what we’re seeing emerge as “agentic RAG” – a concept explored further in [RAG Workflow and Loop Engineering: The Dispatcher That Decides When to Loop and When to Stop]. Das's focus on language-native consumer memory and RQ-VAE semantic IDs highlights a move towards a more nuanced understanding of both user intent and product catalog representation. This isn't just about showing users more of what they've already liked; it's about anticipating their needs and guiding them towards relevant discoveries based on a deeper, semantically-rich understanding of their journey. The grounding of search, tying AI responses to specific catalog items, is a vital practical step in ensuring accuracy and relevance, a point often overlooked in the broader AI hype.

What’s particularly compelling about DoorDash’s approach is its emphasis on building memory and context. Traditional recommendation systems often treat each interaction as isolated, failing to capitalize on the wealth of information accumulated over a user's history. By leveraging language models to encode consumer memory, DoorDash can build a more holistic profile, enabling more personalized and proactive recommendations. This mirrors the exploration of foundational AI agent technologies discussed in [5 Fun Agentic AI Papers to Read], which emphasizes the importance of memory and reasoning capabilities for truly intelligent agents. The choice of RQ-VAE for catalog representation suggests a sophisticated approach to semantic understanding, allowing the system to identify relationships between products that might be missed by simpler methods. Ultimately, this is a move away from reactive recommendations and toward a proactive, conversational experience. The sheer scale of DoorDash’s operation – handling millions of orders daily – makes this a particularly impressive feat of engineering and a strong validation of this architectural direction. Considering the computational demands of these systems, the advancements in AI agent workhorses like [NVIDIA Nemotron 3.5 Lightning] will be essential to support this kind of scaling.

The broader implications of DoorDash’s work extend beyond the food delivery space. The principles of language-native memory, semantic IDs, and grounded search are applicable to any industry dealing with large catalogs and personalized recommendations, from e-commerce to media streaming. The challenge, of course, lies in adapting these techniques to different datasets and business contexts. While the technical details of RQ-VAEs and grounded search might seem complex, the underlying philosophy—treating AI not as a prediction engine but as a helpful agent capable of understanding and responding to user needs—is remarkably straightforward. The success of this approach at DoorDash provides a compelling case study for other companies looking to move beyond legacy recommendation systems and embrace a more agentic approach to AI. The emphasis on practical implementation and measurable improvements in conversion metrics is a refreshing contrast to the often-abstract discussions around AI.

Looking ahead, the question becomes: how will companies balance the sophistication of agentic AI with the need for transparency and control? As these systems become more autonomous, it will be increasingly important to understand *why* they are making certain recommendations and to ensure that they align with business objectives and ethical guidelines. The ability to debug and interpret the decision-making processes of these agentic systems will be a critical skill for data scientists and engineers. Furthermore, the ongoing evolution of large language models and their integration with other AI techniques will undoubtedly continue to shape the future of personalized recommendations, blurring the lines between prediction and conversation even further.