DoorDash details how shared gateways manage AI models and agents
DoorDash has published a technical account of its generative AI platform, whose development began in April 2023. Shared gateways provide model access, track departmental spending and manage agent permissions. The company describes provider fallback when quotas run out, custom human-evaluation interfaces and product-team responsibility for conversation history. Its disclosed usage includes more than 5,000 internal users; the account describes existing infrastructure rather than a new deployment.
Artificial Intelligence··Morning
DoorDash describes model access through one shared gateway
DoorDash, the delivery company, published a technical account of its shared generative AI platform. A model gateway connects teams to different providers through one application interface and software development toolkit, while an agent gateway manages identity, authorization and visibility into agent activity. DoorDash began platform work in April 2023, initially serving machine-learning engineers before expanding access to other engineers and employees outside engineering.[1], [2]
Swaroop Chitlur, who leads the platform, and software engineer Siddharth Kodwani described its operation. Workspaces track consumption and allocate model costs to the department responsible. Quota problems can trigger a move from OpenAI to Azure or from Claude to Bedrock, the model-hosting service. A central interface lets teams inspect requests and responses, including the activity behind their consumption. Teams from the acquired Wolt, Deliveroo and SevenRooms businesses are also being brought onto the infrastructure.[1]
Human evaluation still needs task-specific screens
The engineers said purchased evaluation tools fell short for custom traces, image evaluation and human annotations. For evaluation of a menu image alongside its labels, teams need an interface suited to that job. Product teams work on annotated examples, evaluation criteria and results dashboards, with hands-on support. Conversation history and session management remain product-team responsibilities. Recovering a runtime during long-running work is among the unresolved technical problems.[1]
Tool access separates users, agents and server developers
DoorDash describes users, agents and server developers as separate participants in authorization. Tool-level policies and rate limits can govern write actions. Some safeguards, including redaction of personal information and certain response checks, are optional. The team also describes hosting open-weight models on Modal, a computing service, and using model-serving software vLLM and SGLang. Documentation, help channels and skills plugins support employees learning the infrastructure.[1]
The company reports more than 5,000 internal users, with 45 joining daily, and says 40 percent come from teams outside engineering, including legal, sales, operations and strategy. It also discloses more than 25 agent projects in 2025, more than 50 servers using model context protocol, a standard connecting models with tools, and 300,000 daily tool calls. These are company disclosures of usage, not comparative employee-hours measurements.[1]
Related columns
For more information on this topic, you can read the related columns.
