Event-Driven AI-Vision Inference with Bedrock

Watch the talk

The problem it answers

Data typed in by hand doesn't scale. At two minutes a record, 400 records is 13 hours of error-prone work. The usual fixes are more people, or a dedicated ML platform with ML engineers, SageMaker and an MLOps practice, which takes months before anything reaches a user. For most problems, neither is needed any more. A foundation model is one API call, and the event-driven pipelines your platform team already runs can make it. What makes it safe to rely on is everything around that call.

What you'll take away

By the end, you'll be able to:

• Plan AI into a pipeline you already run as one more step, so it ships through the same reviews, releases and monitoring as the rest of your code.
• Justify checking every model response against a schema before it's stored, so when a model's output drifts, bad data is rejected on arrival rather than found in a report months later.
• Design the pipeline to fan out from storage events, so throughput grows with volume, from 40 images a minute to thousands, with no new infrastructure.
• Choose a model provider without locking yourself in: Bedrock, Vertex AI and Azure AI Vision fill the same slot, and only the SDK changes.
• Select an open format such as Parquet for the results, so the business queries them with the analytics tools it already has.
• Calculate the business case in your own numbers: what manual entry costs at your volume, against a pipeline deployed in days, not months.

See it working

Each pattern is shown working in a live demo: a bookshelf system where a batch of book covers is uploaded, serverless functions process each one as it lands, and a model reads the title, author and ISBN from every cover. I then query the results with SQL, from books by author to the quality of what the model extracted.

Previous and upcoming events

Events where I've run this, or will, with links to each event and its recording where there is one.