Overmind is an open-source model training platform for AI teams. It takes the OpenTelemetry traces your agent already emits and turns them into a smaller specialised model you own, trained and benchmarked with your own evaluators. - Observability: OpenTelemetry-native trace ingest, OTLP straight to the API with no collector, plus a span explorer, sessions and shareable filtered views. Evaluators score root spans as they arrive. - Datasets: curated from production traces, uploads or imports from Langfuse, LangSmith, Braintrust and Galileo. Every row keeps provenance back to its source trace. - Data Workshop: the dataset's own agent prepares and audits the data and proposes steps you approve or deny. Each step is a readable Python cell and a version. - Eval: five evaluator kinds (LLM judge, deterministic, trajectory, statistical, agentic), eval sets and an eval library. The same evaluators run in batch experiments and in continuous production scoring. - Optimiser: rewrites prompts, tool definitions and agent logic as real git diffs and scores them against a baseline. The winning diff is yours to apply. - Training: managed GPU fine-tuning, LoRA or full, of 38 open-weight base models including Qwen, Llama, Gemma and GPT-OSS. Every trained model is benchmarked with your own evaluators, and you can compare it against the model you run in production today. - Inference: OpenAI-compatible serving of trained and frontier models. Weights are downloadable, so the model can run in your own environment. The platform is AGPL-3.0, with an MIT-licensed Python SDK and CLI. Self-host it with Docker Compose, or start free in the hosted console.