Work
Selected applied ML systems and research engineering work in agents, multimodal systems, post-training, and evaluation.
Agents and evaluation
Agentic systems and reliable evaluation
Benchmarks, computer-use agents, and evaluation protocols for foundation models.
DSAgentBench
A benchmark for evaluating whether AI agents can complete end-to-end data-science workflows in real computer environments using notebooks, IDEs, terminals, browsers, and databases.
Evidence: 275 diverse tasks across the data-science lifecycle, 15 evaluated models, and deterministic evaluators. The best reported agent reached 56.70% task success.
Fine-Grained Benchmark Generation
A reference-grounded multi-agent pipeline for generating, verifying, repairing, and deduplicating technically demanding evaluation tasks.
Contribution: Built OCR ingestion and a multi-agent designer and verifier workflow with iterative repair, trace-aware validation, and semantic deduplication.
Evidence: Trace-aware quality control connects source ingestion to final benchmark validation.
DashboardQA
A benchmark for agents that answer questions through real interactive dashboards.
Contribution: Co-created the benchmark to expose failures in grounding, planning, GUI interaction, and multi-view reasoning.
Evidence: Tests filters, tabs, view switching, visual grounding, and cross-view reasoning in realistic dashboard environments.
Large and Tiny LVLM Judges
Automated evaluation protocols for chart models using large and compact vision-language model judges.
Contribution: Evaluated judge reliability, robustness, and efficiency across chart comprehension and reasoning tasks.
Evidence: A 2B judge matched a 7B baseline while reducing inference latency.
Multimodal systems
Data pipelines and visual reasoning
Large-scale multimodal data preparation and rigorous chart understanding benchmarks.
Open-PMC-18M
An open multimodal dataset and scalable pipeline linking biomedical figures with captions and subcaptions for medical vision-language learning.
Contribution: Owned preprocessing and batched vLLM inference for Qwen2.5-VL and Qwen2.5 summary generation.
Evidence: Scales compound-figure extraction, summarization, and OpenCLIP training across 18 million biomedical image-text pairs with optimized batched VLM inference.
ChartQAPro
A challenging chart question-answering benchmark built around diverse real-world visualizations.
Contribution: Co-created the benchmark and evaluation protocol for open and closed vision-language models.
Evidence: 1,341 charts and 1,948 questions evaluated 21 VLMs.
Post-training and generation
Executable and data-driven generation
Feedback-driven post-training and agentic workflows for visualization and narrative generation.
RL-Text2Vis
A GRPO framework that evaluates text-to-visualization outputs after code execution using textual, executable, and visual feedback.
Contribution: Co-developed post-execution multimodal feedback for aligning generated text, code, and rendered visualizations.
Evidence: Improves code executability and chart quality over strong prompting and supervised baselines.
DataNarrative
An Actor-Critic workflow for planning, reflecting on, and revising data-driven stories with visualizations.
Contribution: Introduced a benchmark and multi-agent workflow for automated data-driven storytelling.
Evidence: Includes 1,449 real-world data stories for system and human evaluation.