Without touching model weights: how we built the Alice AI research agent and slashed GPU consumption

Prokhor, lead of the 'Research' agent team for Alice AI, details the evolution of their deep research tool. The agent can create complex plans, perform hundreds of search queries, interact with dynamic web content, and execute Python code for calculations. The article covers the journey from prototype to production, highlighting the product goals set by manager Ruslan Iliev. The focus is on technical optimizations that significantly reduced GPU consumption without modifying the model weights. The authors share lessons learned, including discarding an initial prototype and implementing a multi-layered architecture. This case study demonstrates how a systematic approach to agent development and resource management allows for scaling complex AI solutions for millions of users while maintaining high response quality and system performance.