Variable document formats make structured data extraction difficult to solve with prompting alone. Especially when the documents include everything from images, handwriting, complex financial tables repeated multiple times and without a particular form or naming conventions.

This talk explores how we optimised an intelligent document processing solution on Amazon Bedrock by testing 17 different variants across prompting, model selection, document handling and validation. We will share the methods to optimise an LLM based solution that apply to many use-cases beyond documents, as well as the key trade-offs, failure modes and design decisions that had the greatest impact on extraction accuracy and reliability.

What you'll learn

• How to test and compare AI solution options, including model selection, experiment design and evaluation methods.
• How the findings shaped key design decisions and the final Amazon Bedrock solution released into production.
• Which methods can improve the performance of an LLM-based solution, and the trade-offs between value and implementation effort.

About the Speaker

Silvia Lehnis helps high-impact organisations use data and AI to reach their goals faster and safer. She has led global and national data and AI transformations in sensitive environments across the public sector, finance, energy and academia, taking strategy through to implementation across people, process and technology - with solutions reaching up to 110,000 users and delivering £27m in savings over three years. She's also a board member of the charity Care in Action.