Belvedir

Belvedir is a platform that turns an AI product's production traffic into custom fine tuned models. A single npm package captures every agent call, tool use and reply as data the customer owns; Belvedir then cleans those traces, groups them into recurring task types, builds training datasets, fine tunes an open model on them and scores each run against the customer's own benchmarks. The stated goal is AI that is cheaper, better and faster than calling a general purpose model for every request. It is aimed at teams already running agents in production, such as support automation, that want a smaller specialized model for their recurring tasks. The console reconstructs full conversations from traces, clusters traffic into task groups with success rates and counts of clean examples, and records a cleaning log. Training runs are versioned and reproducible, with evaluation loss and benchmark scores tracked for each model version. Trained models serve on Belvedir's cloud inference behind an OpenAI compatible endpoint, so switching is a one line change, or the weights can be exported and self hosted. Customers need no GPUs of their own, and the FAQ states that traces, data and trained models stay isolated to the customer's workspace and never train anyone else's models. Unlike platforms that start from a finished dataset, Belvedir keeps training as new traffic arrives. Pricing is not published. Users install the package and sign in to create a workspace, and larger teams are invited to contact the founders for onboarding. Belvedir is operated by Fractal Machine Research, Inc.