Belvedir

Belvedir is a platform for training custom AI models and memory systems, hosting them privately, and continually improving them from production data. The target user is a startup already running an AI agent or product on frontier APIs that wants a cheaper, task-specific model without hiring a research team or standing up its own training stack. Founders install an SDK, and Belvedir collects traces from the running agent, converts them into task-specific training sets and RL environments, trains and benchmarks the model, then hosts it locally or in the cloud with a router that picks between the custom model and other endpoints. The launch matters because most teams still default to sending sensitive data to frontier labs since building a private model has meant either paying for forward-deployed engineers or gluing together data pipelines, fine-tuning, evals, and inference on their own. Belvedir bundles those stages into one loop that keeps improving as more traces come in, which is the piece that has kept custom models out of reach for smaller companies. The pitch from the launch video is that setup takes only a few minutes and early customers are seeing meaningful gains on benchmark scores and inference cost, though those numbers come from the company itself. Belvedir is part of Y Combinator's S26 batch and is based in San Francisco. It was cofounded by Zachary Yu and Lance Yan , who previously worked together on Traverse, a training data project aimed at frontier model post-training. For founders and operators tracking where the private model tooling market is going, this launch is worth a look as an early attempt to make the full custom-model workflow a single SDK install rather than a services engagement.