Detecting the moments that matter, frame by frame

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

The proof of concept set out to answer three concrete questions, each a prerequisite for a production build.

The pipeline exposes two Amazon API Gateway endpoints, one for uploading training footage and one for submitting new video for analysis. On the training path, an AWS Lambda function is triggered on upload, chops each video into image frames, and stores them under a category-specific folder in Amazon S3, where they become the labelled dataset used to train the Rekognition model. On the inference path, a second Lambda breaks an incoming video into frames, passes them to the trained model, and reassembles the result so each frame carries the detected event.

The proof of concept was built to a single acceptance test: demonstrate, on the client's own footage, that AWS is a suitable environment for the continued development of an in-game event detection service. It delivered a working end-to-end pipeline, from footage upload through frame creation, model training and per-frame classification, and a honest read on Rekognition's suitability as the foundation for the next stage.

• Frame creation proven — Footage uploaded to AWS is broken into a frame-by-frame breakdown and sorted into category folders in Amazon S3, ready for training.
• Model trained on real data — An Amazon Rekognition model was trained across the seven event categories using the client's labelled dataset, with accuracy reported back honestly rather than assumed.
• Video processed end to end — New video flows through the inference endpoint, is classified frame by frame with a confidence score per event, and returned to the user.
• Reproducible by design — The entire backend is deployed with AWS CloudFormation, so the client can rebuild it and host it inside their own AWS account.
• Handover — A demonstration of the service, supporting materials and documentation were provided through calls and recorded walkthroughs, along with guidance on the steps to reach a MVP and production.

With feasibility evidenced and a working pipeline in hand, the client has a clear, de-risked path from proof of concept to a production event detection service, expanding the training data, refining accuracy per category, and hardening the pipeline for the volume of footage their community produces.