Castform is an RL post-training platform that converts corporate documents into synthetic question-answer pairs to train open-source models on retrieval tasks.

A 4B model post-trained with Castform against Neon's Lakebase Search (BM25 plus vector hybrid) matches GPT-5.6 Sol retrieval accuracy at 100x lower cost per request. Neon provides the infrastructure layer, including dynamic compute scaling for bursty training workloads and database branching for isolated agent rollouts.