DoorDash replaces LLM-only moderation with hybrid AI pipeline to cut latency and cost at scale

DoorDash faced a constraint common to real-time marketplaces: moderating millions of daily messages without disrupting user experience. Pure LLM-based pipelines introduced unacceptable latency (2 to 10 seconds per call) and prohibitive costs at scale. SafeChat’s hybrid architecture addresses this by first applying lightweight internal models to filter clear-cut cases, such as spam or explicit content. This reduces the volume of messages requiring LLM evaluation, cutting both latency and expense. The approach is particularly effective for platforms where interactions are brief and transient, as it avoids over-reliance on slower, more expensive models for routine moderation.

The system’s design prioritizes speed without sacrificing accuracy. By reserving LLM scoring for ambiguous or high-risk content, SafeChat achieves sub-second response times while maintaining safety metrics. This is critical for DoorDash, where interactions between consumers, Dashers, and merchants last only minutes. The no-code workflows and backtesting tools further streamline operations, allowing non-technical teams to adjust moderation rules without engineering overhead. This flexibility is essential for adapting to evolving safety threats or platform policies, though it may introduce complexity in maintaining consistency across rule changes.

SafeChat’s architecture is presented as a repeatable pattern for AI use cases beyond content moderation. The hybrid model, fast filtering followed by targeted LLM evaluation, can be applied to fraud detection, customer support triage, or any scenario where scale and responsiveness are critical. However, the approach has limitations. Internal models require upfront investment in training and labeling data, and their accuracy may degrade if the platform’s user behavior shifts significantly. Additionally, the reliance on LLMs for nuanced decisions means the system remains vulnerable to their inherent biases or inconsistencies, which could lead to false positives or negatives in edge cases.

The presentation highlights a trade-off between cost, speed, and accuracy that engineers must navigate in high-volume AI systems. SafeChat’s success hinges on its ability to balance these factors, but the solution is not one-size-fits-all. Platforms with lower message volumes or less stringent latency requirements might not justify the complexity of a hybrid system. Furthermore, the content-agnostic design, while scalable, may struggle with context-specific moderation challenges, such as sarcasm or cultural nuances. Engineers adopting this pattern should validate its performance against their specific use case and data distribution before full deployment.