Rspamd: A Year of Features and LLM-Assisted Development

This talk covers Rspamd development from March to December 2025, focusing on four major areas. First, HTML fuzzy hashing - a new algorithm that generates structural fingerprints from DOM trees, enabling detection of phishing emails that reuse legitimate templates with modified links. Second, multi-class Bayesian classification that extends the traditional spam/ham model to support up to 20 categories (newsletters, transactional mail, promotions) with single-call Redis lookups. Third, protocol improvements including TCP transport for fuzzy queries and encrypted ZIP archive handling via libarchive. Fourth, neural network refactoring into a provider-based architecture for combining multiple feature sources. We'll also discuss practical experience using LLM tools for code generation, documentation, and PR review during this development cycle - what worked, what didn't, and where human judgment remains essential.