Research

Our research asks a practical question: under what conditions can AI tools be relied on to help people work with documents and records, and what does a person need to see in order to check the result?

Document understanding

Many documents mix narrative text, tables, forms, and references to other documents. We study methods for classifying documents, extracting specific facts from them, and producing summaries that stay faithful to the source. Our current interest is in long documents, where a model must keep track of definitions and cross-references across many pages.

Human review of AI output

AI output is useful only if a person can judge whether it is right. We examine how reviewers check, correct, and come to rely on output from language models, and how software can support that review, for example by showing the source passage behind each extracted fact or by marking statements the system could not confirm.

Evaluation

General benchmarks say little about how a tool will perform on a specific task. We build task-specific test sets and measures of accuracy, consistency, and completeness, and we repeat the same tests when models or methods change so that results can be compared over time.

How we work

We document the data, settings, and methods behind each result so that it can be repeated. We report limitations alongside findings. We use public or synthetic data where possible and handle any other data under the terms on which it was provided.

Inquiries

Researchers and organizations interested in this work may reach us through the Contact page.