Shreya Shankar

Shreya Shankar is giving analysts a way to turn long, unruly documents into inspectable workflows—so that AI can be steered, evaluated and corrected before its answers travel. Her work brings database discipline to systems otherwise built as opaque sequences of prompts.

Much of the world’s consequential data does not arrive as tidy rows. It lives in police records, case files, policy documents, scientific reports and support archives—material too long and irregular for a single model prompt. Shreya Shankar’s work treats this not as a prompting inconvenience, but as a data-systems problem. DocETL, which she led as first author during her doctorate at Berkeley, lets users describe document-processing pipelines in natural language, then uses constrained agents to rewrite, decompose, test and optimise those pipelines for accuracy and cost. Its peer-reviewed 2025 study evaluated four real-world document tasks and reported accuracy gains of 21% to 80% over strong baselines. The software is open source; by July 2026, its public repository had drawn roughly 3,900 stars and remained under active release. At a March 2026 research seminar, Shankar reported broader use across thousands of hosted pipelines and uptake of the project’s ideas by major data platforms. Those figures are signals of interest, not proof of downstream outcomes, and the usage claims remain team-reported. Her larger contribution is conceptual: bringing mature database concerns—query planning, validation, provenance and cost—to AI workflows. She is constructing a practical middle layer between models and the people who must trust their work. The result is not autonomous truth, but a more legible and governable way to make machines work through documents.