Analysis of risk quality metrics for assessing the stability of a large language model

This paper examines the challenge of integrating generative artificial intelligence (LLMs) into corporate systems within the financial sector. The primary difficulty lies in the non-deterministic nature of the models and the risk of hallucinations, rendering traditional software quality metrics and lexical algorithms inapplicable to probabilistic text-based responses.

The aim of the study is to develop a methodology for the continuous monitoring of large language models in production environments. The proposed system integrates a Retrieval-Augmented Generation (RAG) architecture, an "LLM-as-a-Judge" approach, an automated ETL pipeline, and a proactive alerting system to ensure operational business stability.

A "RAG triad" of key metrics was established, comprising answer groundedness and relevance, as well as context accuracy and completeness. A multi-stage PySpark-based algorithm was developed to transform raw dialogue logs into analytical data marts while eliminating key collisions. A dashboard was implemented that demonstrated the statistical significance of the error rate gap between voice (74.32%) and text (16.19%) channels and revealed a correlation between reasoning errors and Average Handle Time (AHT).

The practical value lies in shifting from selective manual checks of 1–3% of communications to automated monitoring of 100% of sessions. This reduces operational risks, frees up to 2,000 man-hours per month by optimizing staff training, and prevents model degradation through a two-level alerting system. Implementation requires a new culture of proactive oversight, transforming AI from an unpredictable tool into a reliable infrastructure component.