The most expensive way to moderate a real-time marketplace is to assume every message needs the same level of scrutiny. DoorDash's Bruna Pereira recently walked through how her team built a content-agnostic AI moderation platform, and the takeaway is a masterclass in cost discipline. Instead of relying on a pure LLM pipeline, which gets pricey and slow when you are processing millions of daily messages, they adopted a hybrid pattern. Fast internal models handle the obvious cases, LLMs step in for nuanced multi-axis scoring, and no-code workflows allow the team to backtest changes before they ever hit production. The result was a measurable drop in safety incidents, achieved while scaling to a volume that would melt a naive architecture.
This is the kind of pragmatic engineering that usually gets overshadowed by headline-grabbing model releases. But it is also where the real transformation happens. We have seen this theme before in our own coverage, like in Exploring Paragraph Structure: How LLMs Navigate Token Space, where the focus is on understanding the mechanics of how models process information rather than just marveling at the output. DoorDash's approach is similar in spirit: it is not about throwing a frontier model at every problem, but about understanding where the model's strengths matter and where a cheaper, faster tool is actually better. And for teams just starting to integrate AI into their workflows, the practical guidance in Unlock ChatGPT for Work: A Practical Guide to Getting Started reinforces this point, that adoption is less about the model and more about the system around it.
What we find most compelling is the discipline around the no-code workflows and backtesting. This is not a flashy feature; it is a governance layer that makes the system auditable and safe. It signals a maturity that is often missing in AI deployments. The lesson here is not that LLMs are overrated, but that they need to be deployed with the same rigor you would apply to any other critical infrastructure. If you are building a safety system for a live product, you need to know why a decision was made, and you need to be able to test a change without taking the whole platform offline. DoorDash built that, and it is a specific, concrete takeaway anyone building a moderation system should steal: start with a fast filter, use the LLM for judgment calls, and always backtest your policy changes.
The open question we are left with is how this pattern generalizes. Is this a template for other real-time systems, or is it specific to moderation? We suspect the former, but the proof will be in how many teams adopt a similar hybrid approach rather than chasing the latest model. The practical takeaway to quote: "Cheap models filter, expensive models judge, and backtesting keeps you honest." That is the architecture of trust in an AI-native world. Watch for how this pattern influences the next wave of tooling, because the team that figures out how to make the LLM the exception, not the rule, will be the one that scales sustainably.