The team led by Alexey, head of the ML department, implements a method to accelerate AI training for anomaly detection. Instead of waiting for historical data, they deploy a pilot system that generates labels while actively operating in real-time. This approach avoids the typical six-to-twelve-month delay required for manual data preparation.
The methodology is applied to procurement control, specifically for identifying inflated prices, non-targeted spending, and off-tender purchases. The team avoids standard supervised learning due to a lack of training data and notes that unsupervised models often produce false positives. Furthermore, LLMs without a reference point struggle to distinguish between market and inflated prices.
The solution involves a multi-layered architecture where a rule-based system serves as the foundation. This layer functions as a data collection engine, gathering necessary labels during live operations. The machine learning model is subsequently integrated on top of this rule-based layer to refine the process and improve detection accuracy over time.
