Belvedir

Teams running AI agents generate large volumes of raw conversation and trace data, but that data is noisy and scattered across tools, making it hard to turn into something usable for training. Most post-training platforms assume a team has already built a clean dataset, which leaves the harder upstream work of collecting, deduplicating, and labeling real usage data unsolved. Without this, teams either keep paying for expensive frontier models on every call or attempt manual dataset curation themselves. The result is that custom models stay out of reach for teams that have the data but not the pipeline to clean and structure it.

Belvedir is installed as an npm package that records every call an agent makes and connects to other apps over MCP to pull in their data, with all of it kept in the user's own workspace. The platform automatically groups this raw data into recurring 'task groups' (for example refund eligibility or order status lookup), cleans it, and tracks how many examples are ready for training. From a task group, a user can fine-tune an open model such as Qwen3.8-27B, with the training run saved alongside its configuration, loss curves, and token counts. Each resulting model is then scored on custom evals against both the base model and a frontier model like gpt-4.1, and can be served through a Belvedir-hosted OpenAI-compatible endpoint or exported as weights to self-host.