tl;dr: Castform uses RL post-training to make small open-source models competitive with frontier models like GPT-5 on agentic retrieval tasks, claiming ~100x cost reduction. The pipeline runs on Neon's Lakebase Postgres, which handles synthetic training data generation, hybrid (BM25 + vector) search during RL rollouts, and autoscaling for bursty training workloads. The pitch: turn your existing corpus into training tasks with a reward function (retrieval + citation + correctness), skipping the usual data engineering and ML infra overhead.