A feature for Amazon Bedrock Data Automation that refines extraction instructions from a few labelled examples, with no model training or fine-tuning. I helped drive its core design and delivery.
Bedrock Data Automation pulls structured fields out of documents. Developers describe each field in a blueprint, in plain language: "the invoice total", "the contract end date".
An instruction that works on one layout often fails on the next. A different label, a subtotal sitting above the real total, a second currency. Developers ended up rewriting instructions by trial and error, which took hours or even days.
An illustration of the loop, not the real service. One instruction, one example document, one round of refinement.
Expected value for total: 248.40
Developers bring up to 10 representative documents from their own workload, each labelled with the values they expect. Bedrock Data Automation compares what it extracts with those labels and rewrites the natural-language instructions in the blueprint until extraction lines up.
My part was helping drive the core design and execution of the feature through to launch. It's rewarding to ship something with such a clear effect on developers' day-to-day work.
Next case study
Amazon Textract async processing