Abstract
<title>Abstract</title> <p>The emergence of Retrieval-Augmented Generation (RAG) has addressed critical limitations of Large Language Models (LLMs), yet standard RAG architectures remain constrained by single-model bottlenecks, static retrieval pipelines and limited capacity for continuous knowledge adaptation in enterprise settings. This paper presents the design, implementation and empirical evaluation of an Automated Learning Engine (ALE) - a multi-LLM orchestrated system that advances beyond conventional RAG towards what we term “runtime intelligence.” The ALE architecture employs a dual-LLM coordination mechanism with deterministic state-machine orchestration, hybrid BM25-vector retrieval at an empirically calibrated 0.55/0.45 weighting ratio and an agentic report generation pipeline that synthesizes four distinct analytical output types. A controlled evaluation was conducted across 847 enterprise queries drawn from financial services domain knowledge bases, comparing ALE against three baseline configurations: standalone LLM, naive RAG and enhanced single-LLM RAG. Results demonstrate that the multi-LLM orchestrated approach achieved a factual accuracy of 94.2% (compared to 71.3% for naive RAG), a contextual relevance score of 0.89 (SD = 0.07) and a 41% reduction in hallucination rate. The hybrid retrieval mechanism contributed a 23% improvement in recall@10 over vector-only search. These findings suggest that moving from static retrieval to runtime intelligence - where multiple specialized LLMs coordinate through state-machine workflows - represents a viable architectural paradigm for enterprise-grade knowledge systems. Implications for both system design theory and practitioner deployment are discussed.</p>