Detecting and Mitigating Hidden Drift Triggers Through Fubini Subset Analysis Traditional data drift monitoring often misses hidden triggers when features interact combinatorially. Just magine, each feature is stable individually, but their joint distribution drifts, this is a "Trojan horse" for models. ➡️ Why Unidimensional Tests Don't Work? Standard tests (PSI, KS, AD) check unidimensional distributions. However, drift can occur only in specific subspaces of features. For example, in a credit scoring model, age and income are stable individually, but among customers with high income and ag

Monitoring and debugging "silent" token drift in LLM pipelines: frequency analysis of embeddings without ground truth An LLM pipeline suddenly starts producing strange responses, even though accuracy doesn't decrease and latency doesn't increase. Standard metrics remain silent, and the responses become increasingly nonsensical. Often, the culprit is "silent" token drift: a subtle shift in the distribution of embeddings without ground truth. The causes can be: updating the model, changing the tokenizer, or the input data gradually shifting. ➜ Spectral analysis of embeddings using FFT One eff

Online detection of feature collisions in TDA transformation When using Topological Data Analysis (TDA) in production, you quickly encounter a problem: persistent diagrams and other TDA features are very sensitive to distribution drift. A new data regime emerges, and the number of topological holes or their significance may not match the baseline. If this is not detected, the downstream model will simply break down, and metrics will start to decline for no apparent reason, such as feature drift. ☞ Detection Architecture The approach is simple. We have streaming data that undergoes a TDA tran