Shreya Shankar

Shankar’s academic work combines data management with the practical difficulties of using imperfect model outputs. Her CV records a Berkeley EECS PhD from 2021 to 2026, a part-time Stanford computer science MS from 2018 to 2020, and a Stanford CS BS from 2015 to 2019. She became a Visiting Assistant Professor at Carnegie Mellon in July 2026; her tenure-track Assistant Professor appointment is scheduled for July 2027.[2]

Changing the plan, not only the prompt

DocETL addresses document tasks that remain difficult even after careful prompting. A model asked to extract every relevant clause from a long document may miss material; the system can instead split the document or decompose the task and combine the results. Shankar and her co-authors define rewrite directives and an agent-guided search that compares alternative processing plans. The paper evaluates the approach on four document-analysis tasks. Its central contribution is making the pipeline itself something the system can optimize for accuracy.[3]

From documents to tables

The open-source DocETL stack lets a user specify operations such as mapping, filtering, and reducing in natural language, then orchestrates model calls over structured or unstructured data. It can return tables for later querying. DocWrangler adds an interactive interface for refining prompts and inspecting results. The repository connects these tools to the research papers and credits the Berkeley groups that developed them, giving users both an implementation and the rationale behind its behavior.[4]

Evaluations begin with the actual errors

With Hamel Husain, Shankar teaches a course on evaluation for people building AI products. It covers inspecting failures, creating task-specific tests, using model judges, and collecting useful human feedback. The course focuses on processes that transfer across vendors and on multi-step systems where a final failure can originate in several different components. It is aimed at people already building an application rather than an introduction to prompting.[5]

Research and teaching beyond DocETL

Her personal site links work on interactive debugging for retrieval-augmented generation, cost-aware semantic data processing, and benchmarks for data agents. It also announces Evals for AI Engineers, an O’Reilly book co-authored with Husain. These projects share a practical concern: helping people inspect, measure, and improve an AI system rather than judging it from a small set of convincing demonstrations.[1]