ML Model and LLM Interpretation for Automating VIP Client Management
01.06.2026
READING TIME 8 MINUTES
Analyzing large datasets to determine VIP client potential often requires significant time resources. As part of the project, a client classification ML model was optimized, reducing its execution time by more than 100 times, and an LLM was integrated to generate clear interpretation of the results. This allowed managers to deliver personalized recommendations in real time.
Context
For effective client service in banking, it is essential to act proactively and offer clients transitions to appropriate privilege levels in a timely manner. To increase the accuracy and speed of decision-making, a need arose to implement an intelligent classification system for analyzing client behavior. gloozd team is extremely attentive to partners' requests not to disclose sensitive data, which is why the bank's name is omitted in this case study.
challenge
Existing scoring models required too much time to process data, making it impossible to use them for instant rendering of personalized screens in the app. Furthermore, standalone machine learning scores (e.g., dry classification scores) did not provide managers with a clear understanding of why a specific VIP status was predicted for a client. As a result, offering personalized recommendations regarding service level changes was complicated.
Fears and how we addressed them
The main concerns were that managers would find it difficult to trust the "black box" decisions of an ML model without understanding the specific contributing factors. We eliminated this barrier by introducing LLM interpretation of the results (powered by SHAP values). The system automatically translates complex mathematical model outputs into simple text explanations for managers, breaking factors down into those that supported the decision and those that lowered the probability.
solution
ML Model Optimization: we optimized the machine learning classification algorithm, cutting data processing time down to under 0.2 seconds. This enabled direct integration of the model into the mobile app to display personalized content. LLM Interpretation: we developed a module based on large language models (LLM) that analyzes key factors (POS transactions, e-commerce spending, geolocation data, connections with other VIP clients) and generates text explanations for managers detailing positive and negative influence factors (SHAP values).
result
Scoring processing and output time was reduced by over 100 times (to less than 0.2 seconds).
Real-time rendering of personalized screens in the mobile app was successfully enabled.
Managers received a convenient assistant tool that explains the rationale behind assigning a status (e.g., PLATINUM), allowing them to offer well-founded and personalized recommendations for transitioning to an appropriate VIP tier.
// ready to start?
