Project Direction

Agentic Data Processing

Emerging agentic AI workflows create a new resilience challenge: natural-language tasks must be decomposed, grounded, executed, and verified over heterogeneous data. My work builds systems that transform such tasks into explicit, traceable, and executable pipelines using structured knowledge representation, hybrid planning, self-correction, and parallel execution — making AI-driven workflows more scalable, auditable, robust under ambiguity, and cost-controlled.

Agentic Data Orchestration for Scholarly Corpora

AgenticScholar: Agentic Data Management with Pipeline Orchestration for Scholarly Corpora (SIGMOD'26) (System Prototype)

Hai Lan, Tingting Wang, Zhifeng Bao, Guoliang Li, Daomin Ji, Ge Lee, Feng Luo, Zi Huang, Hailang Qiu, Gang Hua

AgenticScholar addresses the growing friction in scholarly analysis, where evidence is distributed across modalities such as text, tables, figures, code snippets, and citation and bibliographic context, and where questions naturally evolve from retrieval into multi-step synthesis, comparison, trend tracing, and idea exploration. It treats this interaction as compiling natural-language requests into evidence-grounded analysis workflows represented as explicit executable DAG plans, rather than isolated invocations of a retriever or summarizer. In practice, it supports a spectrum of scholarly queries, including constraint-based paper retrieval; single-paper extraction and structured summarization of problem, method, and experimental details; cross-paper synthesis with ranking and inconsistency checking; and higher-level knowledge discovery such as research trend analysis, milestone paper selection, and uncovering under-explored problem–method combinations. Its agentic core unifies a structure-aware, multi-modal scholarly knowledge base with hybrid planning and operator-based execution, while keeping the process inspectable via materialized plans, intermediate results, and data lineage, and supporting new workflows by composing and extending a library of reusable operators.

AgenticScholar system overview