Building Jarvis Pro: Route first, answer later

This post details the evolution of Jarvis Pro's architecture with a 'route first, answer later' design. Key technical contributions include the development of a prompted LLM classification router that determines task type, scope, and necessary context, preventing common errors by loading only relevant information. The system's memory management was refined to keep context narrow and prioritize explicit context over saved defaults. A deliberate tradeoff was made between first token latency and response safety, with slower first tokens being acceptable for safer recommendations. Regression checks were implemented to normalize source values and compare daily rows across approved metric paths for metric reconciliation. Evaluation was conducted using a routing set and an answer-quality suite, focusing on accepted route accuracy and rubric-based scoring.