Adaptive LLM Router
Deployed an adaptive multi-LLM routing system, ensuring precise agent state management while dynamically switching models at tool call boundaries across 500+ SWE-Bench Verified software engineering tasks. Created a backend system with modular model adapters, parallel task execution, human-in-the-loop routing, session management, persistent turn history, and real-time tracking of token usage and tool calls. Integrated local Qwen models via vLLM/SGLang along with frontier OpenAI Codex models through an adapter interface, enabling cross provider model switching and context preservation within active agent tasks. Refined automated routing signals from testing data including runtime failures, token usage, dependency errors, and execution progress, reducing token costs by 36% while maintaining task success rate over single model baselines.