Transformers’ training and export interfaces suggest a further distinction that is relevant to DVSA API: DVSA does not need to provide every training capability in order to define a training contract. Such a contract can specify how aerial datasets are registered and loaded; how labels, temporal annotations, and geospatial features are represented; which metadata must accompany a trained model; which export targets are accepted, including ONNX, TensorRT, or Azure Custom Vision variants; and how an export is validated against the DVSA inference adapter. Under that arrangement, a model may be trained in any compatible toolchain, exported with the required metadata, installed into DVSA API, and made available to pipeline execution, MCP-driven workflows, and retrieval-based queries.

The same abstraction can extend beyond individual inference models. DVSA API’s Multi-Agent Control Plane and pluggable MCP support can represent agents by capabilities and contracts, workflows as graphs connecting those agents, and mission templates as versioned packages. A search-and-rescue package, for example, could declare object detectors, aerial reasoning models, flight-path planners, sensor-actuator interactions, storage targets, and human escalation rules as one coherent workflow. This is comparable at a structural level to combining a model pipeline with a reusable application, while remaining specific to aerial sensing and operational control.

For end-users, the mapping can therefore be read in operational terms. The pipeline configuration states what should run and how the stages are connected; detector and reasoner packages provide task-specific inference; preprocessors normalize video, imagery, metadata, and geospatial context; post-processing converts raw predictions into DVSA records; storage adapters persist those records; and MCP workflows coordinate subsequent analysis, planning, or escalation. A registry and associated cards make these components discoverable, while versioning and validation establish whether a package can be used with a given DVSA API release.

Implementing this model would involve formal, versioned contracts for configurations, detectors, reasoners, and preprocessors; a high-level pipeline API; manifests and cards for models and complete pipelines; a dataset registry designed for aerial video and geospatial annotations; a training and export contract for plug-in inference; and a shareable format for MCP agents, workflows, and mission templates. Together, these elements would apply the parts of the Transformers approach that are relevant to DVSA API: consistent interfaces, explicit metadata, portable artifacts, and task-level composition across otherwise different models and runtimes.

References

1. DVSA API repository: https://github.com/ravibeta/dvsa-api