Scaling computer vision inference on AWS

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

The evaluation focused on three questions that decide whether a machine learning pipeline can scale sustainably.

Video is uploaded to an Amazon S3 input bucket. That upload triggers an AWS Lambda function, which submits a new job to an AWS Batch job queue. The job runs the YOLO model inference on the AWS Batch compute environment using a container image held in Amazon ECR, and writes its results, the bounding boxes and predictions, to a S3 output bucket.

The project was accepted against its success criteria: a feasibility assessment of AWS Batch for YOLO inference, validated instance selection, a cost-aware environment with auto scaling, and documentation and knowledge transfer for a clean handover, all deployed into the client's AWS environment via CloudFormation. The figures below are drawn from the Statement of Work's scaling model and are projections.

• A proven scaling engine — AWS Batch was validated as a practical way to manage and scale YOLO inference jobs, with job queues, retry strategies, and managed compute removing server-management overhead.
• An evidence-based compute choice — Benchmarking CPU Fargate against GPU EC2 gave the client a quantified basis for choosing instance types on performance, cost, and resource utilisation rather than guesswork.
• Cost that follows demand — Auto scaling on job-queue depth means capacity rises under load and falls when idle, keeping the solution cost-effective as volumes grow toward tens of thousands of videos per day.
• Event-driven and hands-off — A S3 upload automatically triggers a Lambda that submits the Batch job, so inference runs without manual intervention.
• Handover — Calls, Loom walkthroughs, and documentation were provided so the client can operate the pipeline and take it toward production.

The pipeline delivers the start of the client's ML inference capability on AWS, a foundation that is no longer a proof of concept but a scalable base to build on as demand grows and the service moves toward production.