Hello! I am Bhavya Chopra, a third-year CS PhD student at the EPIC Data Lab at UC BerkeleyUC Berkeley, advised by Prof. Aditya Parameswaran. My research makes AI-powered data systems steerable through shared representations that users and agents can inspect and directly manipulate, and algorithms that persist and enforce user feedback at scale.

We are building an interface for GEPA that lets you observe and directly steer long-running AI-powered optimizations. Join our study for early access and to help shape the interface!

I recently interned at TableauTableau Research, where I was thrilled to be mentored by Arjun Srinivasan. Previously, I was a Research Fellow with the MicrosoftPROSE Team at Microsoft, advised by Austin Z. Henley, Chris Parnin, and Anna Fariha. My work focused on identifying challenges in conversational and recommendation-based AI systems for data science and exploring agent-based frameworks to enable collaborative debugging with large language models.

Before that, I completed my bachelor's in Computer Science and Design engineering (with honors) from IIIT-DelhiIIIT-Delhi in 2022. I developed interfaces for inclusive and immersive storytelling experiences for children with visual impairments, and also learnt program analysis techniques to study the use of assertions in Java open-source repositories.

Updates

Publications

HoldUp workflow: unlabeled records are linked in a graph, grouped by correlation clustering, and each cluster is matched to one of the class names Sports, Finance, Technology and World by bipartite matching.
Semantic Data Processing with Holistic Data Understanding
Youran Sun*, Sepanta Zeighami*, Bhavya Chopra, Shreya Shankar, Aditya G. Parameswaran
To appear at VLDB 2027
Who's Keeping Score? Interactive Steering of LLM-Powered Scoring with Attune
Bhavya Chopra, Meng Chen, Rebecca Dang, Chanbin Park, Shreya Shankar, Sepanta Zeighami, Bjoern Hartmann, Aditya G. Parameswaran
To appear at UIST 2026 • Try Attune
The MAST taxonomy of multi-agent system failure modes: 14 failure modes across pre-execution, execution and post-execution stages, grouped into system design issues (44.2%), inter-agent misalignment (32.3%) and task verification (23.5%).
Why do Multi-Agent LLM Systems Fail?
Mert Cemri*, Melissa Z Pan*, Shuyi Yang*, Lakshya A Agrawal, Bhavya Chopra, Rishabh Tiwari, Kurt Keutzer, Aditya G. Parameswaran, Dan Klein, Kannan Ramchandran, Matei Zaharia, Joseph E. Gonzalez, Ion Stoica
NeurIPS 2025 Datasets and Benchmarks Track • Spotlight Paper
The three gulfs of semantic data processing between the user, the data and the LLM pipeline (comprehension, specification and generalization), and the DocWrangler features that address each. The DocWrangler IDE: a notebook-style pipeline editor, an output table with per-column histograms, and a dataset panel
🏆 Steering Semantic Data Processing with DocWrangler
Shreya Shankar*, Bhavya Chopra*, Mawil Hasan, Stephen Lee, Bjoern Hartmann, Joseph M. Hellerstein, Aditya G. Parameswaran, Eugene Wu
UIST 2025 • Best Paper Honorable Mention • Read our blog post • Try DocWrangler
The DataScout interface: query reformulation suggestions and semantic filters on the left, ranked dataset results in the middle, and a dataset detail view with relevance indicators and a preview table.
Rethinking Dataset Discovery with DataScout
Rachel Lin*, Bhavya Chopra*, Wenjing Lin, Shreya Shankar, Madelon Hulsebos, Aditya G. Parameswaran
UIST 2025
Distribution of 114 data scientists' agree and disagree survey responses, for example 92% would go back and forth with ChatGPT to refine a prompt and 87% expect generated code to need several changes.
Challenges in Using Conversational AI for Data Science
Bhavya Chopra, Ananya Singha, Anna Fariha, Sumit Gulwani, Chris Parnin, Ashish Tiwari, Austin Z. Henley
Workshop on Human-In-the-Loop Data Analytics (HILDA), SIGMOD 2025 • featured on Austin's blog
Timelines of 16 participants debugging two tasks with the baseline assistant and with Robin, showing time spent on understanding, localization and fixing, and where localization and fixes succeeded.
🏆 Let's Fix This Together: Conversational Debugging with GitHub Copilot
Yasharth Bajpai, Bhavya Chopra, Param Biyani, Cagri Aslan, Dustin Coleman, Sumit Gulwani, Chris Parnin, Arjun Radhakrishna, Gustavo Soares
VL/HCC 2024 • Best Paper Award • featured on Microsoft Research Focus Blog and Microsoft New Future of Work Report 2024
From a data table, a language model generates candidate question and code pairs. One is dropped at the Executes? check because its code fails to run, one is dropped at the Aligned? check because its code does not answer its question, and the remaining pairs become green Insights.
🏆 Semantically Aligned Question and Code Generation for Automated Insight Generation
Ananya Singha, Bhavya Chopra, Anirudh Khatry, Sumit Gulwani, Austin Z. Henley, Vu Le, Chris Parnin, Mukul Singh, Gust Verbruggen
LLM4Code Workshop, ICSE 2024 • Best Paper Award • featured on Microsoft Research Focus Blog
Detangler's pipeline view of a data wrangling chain, with row and column counts beside each step and boxes colored by change type. Table view with the newly changed column highlighted for the selected step Data Details view with missing-value bars, column histograms and potential-problem counts
🏆 Detangler: Helping Data Scientists Explore, Understand, and Debug Data Wrangling Pipelines
Nischal Shrestha, Bhavya Chopra, Austin Z. Henley, Chris Parnin
VL/HCC 2023 • Best Paper Award
CoWrangler's ranked list of data wrangling suggestions with explanations, such as splitting the title column on a colon. Dataframe preview of the 911 calls data with the changed columns highlighted The pandas code CoWrangler generates for the accepted suggestion
CoWrangler: Recommender System for Data Wrangling Scripts
Bhavya Chopra, Anna Fariha, Sumit Gulwani, Austin Z. Henley, Daniel Perelman, Mohammad Raza, Sherry Shi, Danny Simmons, Ashish Tiwari
SIGMOD 2023 Demonstration
The StoryBox prototype: a yellow wooden box with tactile clay paintings of a monkey and crocodiles, an interactive mat with braille-numbered buttons, and a tray of clay figurines. Close-up of a tactile clay painting of the monkey riding on the crocodile's back The open StoryBox seen from above, with tactile painting panels and clay figurines stored inside The tray of tactile clay figurines
StoryBox: Independent Multi-modal Interactive Storytelling for Children with Visual Impairment
Bhavya Chopra, Richa Gupta
CHI 2022 Late-Breaking Work
Reality Tales home screen with the tagline 'Hear the difference' and a banner inviting readers to talk to the story characters. Voice call screen for conversing with a story character A participant talking to a story character in Reality Tales on a laptop A participant listening to a story in Reality Tales on a phone with headphones
🏆 Reality Tales: Facilitating User-Character Interaction with Immersive Storytelling
Bhavya Chopra*, Khushali Verma*, Sonali Singhal*, Utsav Singla*
CHI 2021 Student Research Competition • Second Position (undergraduate category)

Academic Service

Head Teaching Assistant at UC Berkeley

Data Engineering (Data 101 / CS 187, Fall 2026)

Teaching Assistant at IIIT-Delhi

Program Analysis (CSE503, Monsoon 2021), Introduction to Human-Computer Interaction (DES102, Winter 2022)

Reviewing

UIST (2025, 2026), CHI (2023, 2026), CHI Late-Breaking Work (2022, 2024)

Student Volunteering

CSCW (2021, 2022), CHI PLAY 2022, IndiaHCI 2021

Selected Software Projects

Plants versus Zombies: Strategy Video Game

Re-implemented Plants vs Zombies in Java, following object-oriented programming principles.

GitHub Video

TechTree: Course Repository for IIIT-Delhi

Designed, developed, and hosted TechTree for the academic department and faculty at IIIT-Delhi to maintain and share descriptions of over 350 courses with students. TechTree enables 3000+ students every semester to plan their courses through visualization of course pre-requisites via interactive tables and directed acyclic graphs.

Course Directory Course Dependency Graph