{"id":177,"date":"2026-10-03T16:53:38","date_gmt":"2026-10-03T13:53:38","guid":{"rendered":"https:\/\/faidaworkforce.com\/blog\/the-doordash-playbook-scaling-genai-for-5-000-users-without-losing-con\/"},"modified":"2026-10-03T16:53:40","modified_gmt":"2026-10-03T13:53:40","slug":"the-doordash-playbook-scaling-genai-for-5-000-users-without-losing-con","status":"publish","type":"post","link":"https:\/\/faidaworkforce.com\/blog\/the-doordash-playbook-scaling-genai-for-5-000-users-without-losing-con\/","title":{"rendered":"The DoorDash Playbook: Scaling GenAI for 5,000 Users Without Losing Control"},"content":{"rendered":"

From Chatbot Experiments to Enterprise Infrastructure<\/h2>\n

When OpenAI launched ChatGPT in late 2022, the immediate reaction across many technology organizations was a rush to sign contracts and integrate external models. For DoorDash, this moment arrived roughly four months later, in April 2023. The initial impulse was simple: start using OpenAI. However, the leadership team quickly recognized that a vendor-first approach would not scale. As Swaroop Chitlur, who leads the Generative AI Platform at DoorDash, noted, the concern was not just about the immediate cost, but about how to support such a large company productively and accountably if the technology took off.<\/p>\n

This realization marked the beginning of a deliberate journey to build an internal GenAI platform. The goal was not to create another chatbot or a coding assistant, but to provide infrastructure that would accelerate development velocity and improve production reliability for over 5,000 internal users. The story of DoorDash\u2019s GenAI platform offers a clear blueprint for businesses navigating the transition from experimental AI to embedded operational tools.<\/p>\n

Defining Principles Before Building Tools<\/h2>\n

One of the most critical steps DoorDash took was establishing operating principles before writing significant code. The team was given a “blank slate” and had to decide how to operate. They adopted tenets that prioritized being customer-obsessed and focusing on business impact rather than building systems for their own sake. A key insight was that their primary customer was not the machine learning engineer, but the product engineer.<\/p>\n

In early 2023, the team realized they were no longer talking exclusively to ML engineers who understood notebooks and GPUs. Their audience expanded to include all engineers, and eventually, non-engineers. This shift required a fundamental change in strategy. Instead of providing raw compute access or complex notebook environments, the platform needed to be API-first. The ethos became one of building products, not just systems, ensuring that best practices were baked into the platform\u2019s operation so that making the right choice was also the easy choice.<\/p>\n

The team explicitly decided to focus on business impact. They identified three broad buckets of value where GenAI could help product teams: automation, recommendations, and personalization. Automation was seen as a way to decrease the bottom line (cost), while recommendations and personalization were viewed as ways to increase the top line (revenue). The unique selling proposition of the platform became helping these product teams optimize the trade-offs between accuracy, latency, and cost.<\/p>\n

The Challenge of Fragmentation and Velocity<\/h2>\n

As adoption grew, the technical landscape became increasingly fragmented. Product teams wanted to experiment with different providers\u2014OpenAI, Claude, Gemini, and various open-source models. Each team attempting to build its own plumbing for each provider resulted in slowed development speeds. The frequency of new model releases meant that teams needed to test whether a new open-source model performed better than an existing proprietary one, but the integration overhead was prohibitive.<\/p>\n

To solve this, DoorDash developed an LLM Gateway. This concept provided a single API and SDK interface for the entire company. By abstracting the underlying provider differences, the platform allowed teams to make one change to switch models. This approach significantly improved velocity, enabling rapid experimentation without reinventing the integration wheel for every new model or provider.<\/p>\n

However, speed alone was not enough. The team encountered significant capacity issues. Even when teams wanted to use specific models like Claude or OpenAI, they faced rate limits and quota shortages. For production services, relying on external vendors without robust fallback mechanisms was risky. If a vendor ran out of resources, the application could not function. This necessitated a deeper understanding of routing and fallback strategies to ensure reliability.<\/p>\n

Adapting to a Non-Technical User Base<\/h2>\n

A surprising outcome of DoorDash\u2019s platform design was its adoption by non-engineers. While the initial focus was on engineering teams, approximately 40% of the platform\u2019s users are now non-engineers, including staff from legal, sales, operations, and strategy. These users interact with the platform via APIs, leveraging the simplified interfaces built by the GenAI team.<\/p>\n

This expansion of the user base highlighted the need for accountability and reliability as the surface area of the platform grew. The team\u2019s original belief in building for velocity remained central, but it had to be balanced with rigorous standards for reliability. As the platform enabled more stakeholders, the importance of end-to-end workflows, onboarding, and support increased. The platform had to think like a product team, ensuring that users from diverse backgrounds could successfully onboard and derive value without needing deep technical expertise in AI infrastructure.<\/p>\n

Lessons for Business Leaders<\/h2>\n

DoorDash\u2019s experience underscores several actionable lessons for organizations considering their own GenAI strategies:<\/p>\n