Most companies possess valuable internal data—such as documentation, support articles, and operational databases—but struggle to use it effectively for training AI agents due to the complexity of data preparation and infrastructure demands. According to Ying Hang Seah, cofounder of Castform, pointing Castform at Neon skips both the challenge of turning raw data into usable training material and the need for advanced infrastructure to support scalable, low-cost agent operations. Castform addresses these barriers by converting existing corporate knowledge bases into synthetic training tasks and managing the reinforcement learning (RL) post-training process that teaches open-source models to retrieve and reason over that data effectively. The system requires three components: a task (e.g., answering a user’s question), an environment (a search tool over the corpus), and a reward function (e.g., measuring answer correctness). With these in place, RL post-training proceeds via trial and error, where the model improves through feedback on its search and reasoning attempts. During training, agents repeatedly call Neon’s Lakebase Search to retrieve relevant context, generating a bursty workload that Neon handles via dynamic compute scaling—automatically increasing resources during demand spikes and scaling to zero during idle periods. This avoids the need for over-provisioning and reduces costs significantly. Furthermore, Neon’s branching and time-travel features allow each training rollout to operate in an isolated, resettable database environment, preventing interference between parallel experiments and eliminating the need to maintain thousands of always-on compute instances. As noted in the source, a typical multi-turn search request using gpt-5.6-sol takes over 10 seconds and costs approximately $0.03 end-to-end, making it slow and expensive for production use. In contrast, RL post-trained open-weights models, while initially less capable, can match or surpass frontier models on specific tasks like retrieval when properly trained—achieving this performance at orders of magnitude lower cost per request. Castform streamlines this process by removing the need for deep machine learning or GPU expertise, making post-training as accessible as prompt engineering. The platform also provides observability tools to monitor reward progression and inspect individual model behaviors during training, helping detect issues like reward hacking or tool failures. By enabling developers to turn their proprietary data into high-performing, low-cost agents, Castform and Neon together offer a path toward affordable, scalable agentic search that outperforms expensive frontier models without requiring prohibitive infrastructure investment.
Key facts
- A typical multi-turn search request with gpt-5.6-sol takes >10s and costs ~$0.03 end-to-end
- Small open-weights models are 100x cheaper than frontier models but lag in capabilities out of the box
- RL post-training helps bridge the capability gap, allowing open-source models to match or beat frontier models on specific tasks like search
- Castform enables RL post-training without requiring expertise in machine learning or GPU internals
- Castform’s pipeline uses Neon Lakebase Search for corpus storage, synthetic data generation, RL training, and production inference
- Neon’s dynamic compute scaling absorbs bursty workloads during training, providing low-latency search when needed and scaling to zero during idle periods
- Neon branching allows isolated database states for each training rollout, preventing interference between parallel experiments
- Time-travel queries in Neon enable reconstruction and inspection of agent states during training
- Castform turns existing corporate data (e.g., internal docs, support articles) into synthetic question-answer training datasets
- The reward function in Castform training specifies desired model behaviors, such as retrieving correct chunks and citing sources properly