DoorDash AI Research Optimizes Code Review with Hybrid Models

DoorDash's artificial intelligence research division has developed an innovative method for enhancing its internal code review processes. They created a benchmark, dubbed DashBench, which re-evaluated 105 historical code changes made by their engineers. This system uses AI models to identify and analyze potential issues within new code submissions. The core of their discovery involves a two-stage AI review process. Initially, a model quickly scans code changes for suspicious areas, and then a second, more powerful model performs a detailed examination of those flagged sections. While initially using Anthropic models for both stages, their research revealed that a more cost-effective and efficient approach involves using an open-source Chinese model, Kimi K2.6, for the initial skimming, paired with Anthropic's Claude Fable 5 for the in-depth analysis. This hybrid setup demonstrated superior performance, catching 65.2% of real problems, including 8 out of 10 critical bugs, compared to 53.6% for the purely Anthropic system. Furthermore, it achieved a slight cost reduction, dropping from $3.91 to $3.81 per code change. DoorDash co-founder Andy Fang highlighted that this internal testing framework is crucial for integrating open models, leading to both better quality and lower operational expenses.