A high-performance, scalable text processing pipeline for eDiscovery is outlined. The classification module of the pipeline is based on the random forest model which is fast, exible and allows for relevance scoring and feature importance coupled with high-accuracy results. The feature selection approach combines natural language processing with legal domain input, and is based on regular expressions, which allows for linguistic variation and subtle ne-tuning. These two components of the pipeline are described in some detail. Briefly discussed are a number of the other features, which include relevance hypothesis testing, deduping and social communication network analysis.