Castform has developed a platform designed to enable Reinforcement Learning (RL) post-training for open-source models, allowing them to match or outperform frontier models like gpt-5.6-sol in agentic retrieval tasks at a 100x lower cost. While typical multi-turn search requests using gpt-5.6-sol incur latencies exceeding 10 seconds and costs of approximately $0.03 per request, Castform leverages small open-weights models to drastically reduce these figures. The system achieves this by automating the transition from raw enterprise data—such as internal documentation and operational databases stored in Neon's Lakebase Postgres—into structured training tasks. The pipeline utilizes Lakebase Search, specifically the lakebase_text and lakebase_vector extensions, to facilitate synthetic data generation and tool-call rollouts during the RL loop. By integrating with Neon's infrastructure, Castform benefits from dynamic compute scaling to handle bursty search workloads and utilizes database branching and time-travel queries to provide isolated, reproducible environments for stateful agent rollouts. This approach enables developers to build specialized agents that can effectively navigate, search, and mutate proprietary datasets through iterative trial-and-error training guided by custom reward functions.