python-api.md

Python API

AutonomyFit remains CLI-first, with a small Python API for embedding model recommendation and safe deployment assessment in other software.

Recommend models

from autonomyfit import recommend

recommendations = recommend(
    task="detection",
    hardware_profile="nvidia-t4-16gb",
    objective="latency",
    offline=True,
    limit=3,
)

for item in recommendations:
    print(item.model.id, item.verdict, item.confidence.score if item.confidence else None)

Omit hardware_profile to assess the current machine. The function uses the same evidence, constraint, confidence, Pareto-ranking, and registry logic as the CLI recommendation path.

Common constraints such as latency, throughput, power, accuracy, memory, parameter count, runtime, precision, and minimum confidence can be supplied as keyword arguments.

Assess deployment compatibility

Assess a model before selecting an artifact:

from autonomyfit import assess_deployment

assessment = assess_deployment(
    "yolo26n",
    hardware_profile="nvidia-t4-16gb",
    offline=True,
)

print(assessment["status"])

Or inspect a local artifact:

assessment = assess_deployment(
    "yolo26n",
    artifact="./yolo26n.onnx",
    runtime="onnx",
    expected_sha256="EXPECTED_SHA256",
)

assess_deployment() uses the existing deployment validator and returns its structured assessment dictionary.

Deliberate API boundary

The small Python API does not expose remote artifact acquisition, conversion, or benchmarking. Those workflows have larger trust, execution, and reproducibility surfaces and remain available through the autonomyfit CLI.

The public package surface is intentionally limited to:

• recommend()
• assess_deployment()
• DeploymentValidationError

Internal modules remain implementation details while AutonomyFit is pre-1.0.