overbae backend map

Single Django app overbae, project-scoped tenancy.

Layout

• api/ — DRF views and serializers, one module per surface (eval_views.py, datasets.py, optimizer.py, otlp.py, billing.py, mcp.py, …). views.py and serializers.py are legacy monoliths; a new surface gets its own module. The global exception handler (api/exception_handler.py) flattens list errors to {detail}, adds code, maps IntegrityError to 409 conflict, and returns {detail, code: internal_error, error_id} for unhandled errors.
• models/ — split by domain (iam.py, capabilities.py, traces.py, innermind.py, finetuning.py, datasets.py, dataset_pipeline.py, operations.py, billing.py, …).
• innermind/ — the eval matrix evaluators, live span and trajectory scoring, deterministic checks, evaluation runs, the Default eval set, health and signals.
• services/ — business logic. Largest subtrees: datasets/ (Data Workshop; data-workshop skill), mcp/ (mcp skill), connectors/, sft_assets/ (training code baked into the Modal image), codebase/.
• tasks/ — Celery tasks, roughly one module per feature; utils/task_lock.py for locking.
• modal/ — Modal app entrypoints (vLLM serving, SFT, native decision evaluation and serving) and models.json. Code shared by those workers and the API lives in the root modal_shared/ package.
• management/commands/ — seed_demo (seed-demo-data skill), grant_credits, check_workshop_transition, sync_baseten_finetuning_costs.

Patterns every subsystem follows

• One domain service, three surfaces. REST, MCP and the Console call the same service. Never copy business logic into a view or an MCP adapter.
• Request keys. An operation that starts work takes a project-scoped request_key. An identical retry returns the original record; a changed payload under the same key conflicts.
• Durable admission. Waiting work lives in PostgreSQL rows, and a locked dispatcher admits it with per-project fairness. Do not park waiting work in the broker.
• Unknown provider outcomes are reconciled, never resubmitted. Intent is journaled before a provider call; a lost acknowledgement becomes submission_unknown and is resolved from the saved call ID.
• Measured, not invented. Missing usage, progress, scores or telemetry stay unknown or null. Never fill a percentage, a zero score, an ETA or a cost from elapsed time or a guess.
• Status writes use .filter().update(). No Django signals fire; reload the row and wire side effects at the call site.
• Model IDs: platform models (judges, generation, Jev) live in core/model_registry.py; the training and serving catalog lives in modal/models.json, read through modal/model_registry.py. Jev decisions go through core/decisions.py.
• Modal code ships separately. Changes under overbae/modal/, modal_shared/ or services/sft_assets/ reach running jobs only after modal deploy; training and serving pin immutable releases.

Reference files

Read the one that matches the task: