A reproducible computer-vision platform on AWS

A SaaS B2B business working with Cloud Combinator on AWS. The client is anonymised at their request.

The engagement had four focus areas, each captured as a concrete success criterion.

The work was structured as three phases across roughly twelve weeks, each with its own acceptance criteria, taking the platform from an usable minimum through reproducible pipelines to a fully operated and cost-controlled service.

The platform was accepted against its success criteria: an image submitted through the API returns asynchronous outputs in S3 tagged with the model version and dataset manifest, retraining from a manifest reproduces equivalent metrics, and only approved models reach production. The client moved from two locally run experiments to an operated, auditable service. The figures below are the design targets the platform was built and validated against.

• One cost-aware API — Serves both the turf and weed models through asynchronous SageMaker inference, with preprocessing inside the containers so there is no training-to-inference drift.
• Reproducible, versioned pipelines — Record the model package, dataset manifest, container digest and code revision for every run, so results are auditable and retraining is deterministic.
• Cost efficiency by design — Uses Spot training, asynchronous and batch inference and S3 lifecycle policies to Glacier IR, keeping cost per thousand images inside an agreed envelope.
• Operable and monitored — Gives the team CloudWatch dashboards and alarms plus SageMaker Model Monitor drift detection, so issues are diagnosable from the console and drift can trigger retraining.
• Handover — Delivered the API contract, a golden test set, a "add a new model" guide, incident runbooks and infrastructure-as-code so the client can extend the platform independently.

With the foundation in place, the client can add new models through the documented path, lean on drift-triggered retraining to keep quality high, and scale comfortably beyond today's throughput as its imagery and customer base grow.