OpenProcessor System Overview

OpenProcessor System Overview An architecture diagram generated by Archify. Clients · curl / Cropwright · Architecture component Clients curl / Cropwright Monitoring · Prometheus / Grafana · Architecture component · profile: monitoring Monitoring Prometheus / Grafana profile: monitoring api · FastAPI :4603 · Architecture component · core + /curation api FastAPI :4603 core + /curation triton-server · TensorRT :4600-4602 · Architecture component · GPU triton-server TensorRT :4600-4602 GPU OpenSearch · k-NN :4607 · Architecture component OpenSearch k-NN :4607 Job volumes · /jobs + state dir · Architecture component Job volumes /jobs + state dir Curation workers · all active projects · Architecture component · profile: curation Curation workers all active projects profile: curation segmenter · SAM 3 :4611 · Architecture component · profile: segmenter segmenter SAM 3 :4611 profile: segmenter VLM endpoints · per-project choice · Architecture component · profile: vlm or BYO VLM endpoints per-project choice profile: vlm or BYO curation-trainer · Ultralytics · Architecture component · profile: training curation-trainer Ultralytics profile: training curation-evaluator · bake-off harness · Architecture component · profile: curation curation-evaluator bake-off harness profile: curation MLflow · :4609 · Architecture component · profile: training MLflow :4609 profile: training openprocessor CLI · vlm use / key set · Architecture component · host openprocessor CLI vlm use / key set host HTTP gRPC infer + load index + k-NN + config job.json scrape items per project region models segment active endpoint claim job bake-off jobs metrics vlm select + probe Legend Backend Database Cloud External

Core (always on)

  • • api, triton-server and opensearch are the only services without a compose profile
  • • One shared Triton gRPC client lets Triton's dynamic batcher form batches
  • • No authentication on any route: run behind your own proxy

Opt-in profiles

  • • curation: detection, VLM, auto-label and cluster-refresh workers plus the evaluator, shared by all projects
  • • segmenter, vlm and training each start their own GPU service
  • • The API also calls the VLM directly for /curation/projects/{project}/vlm/* routes

File protocols

  • • Training: job.json in, status.json / run.log / manifest.json out, per project
  • • Bake-off jobs are files the evaluator watches, per project
  • • Promote writes the model repository as {slug}__name and asks Triton to load it