codebreaker77/x-algo-breakdown

X Algo Breakdown is a documentation repository containing a technical analysis of the X "For You" recommendation algorithm, based on the X source code released on May 15, 2026. The analysis is grounded directly in the Rust and Python source files, tracing execution paths through the actual codebase. The breakdown covers four primary subsystems: the Phoenix ML Engine (a two-tower retrieval system with a transformer-based ranking model that uses a candidate isolation attention mask to score posts against a user's history), the Rust Candidate Pipeline (a concurrent pipeline that hydrates, scores, and filters candidates at runtime), the Home Mixer (a gRPC service that coordinates over 28 external service clients, manages feature flags, blends organic posts with ads via the SafeGap blender, and produces the final ranked feed), and the Grox AI Daemon (a Kafka driven Python engine that runs Vision-Language Models to detect policy violations, assign quality scores, and extract embeddings from posts and videos). The scoring model uses 19 distinct engagement prediction weights routed through the Grok-1 transformer. Post storage uses an in-memory system called Thunder for sub-millisecond access. The codebase analyzed spans 207 source files, 139 in Rust and 68 in Python, organized into ten chapters. The repository also includes algorithmic playbooks derived from the source code's scoring multipliers, translated into content strategy guidance, plus notes on what changed since the 2023 algorithm release. It is intended for systems engineers, ML researchers, and product builders. The original X source code is licensed under Apache 2.0, this analysis document itself is provided for educational and research purposes.