REVA: Reusable Evidence View Aggregation for Context-Efficient RAG Serving
Undergraduate Research Assistant, 2026.02-2026.09, Champaign, IL
- Proposed attention-based scoring for post-retrieval RAG context compression to reuse the generator's own attention traces across queries with a document-keyed score store built offline, achieving a 92.0% cache hit rate on HotpotQA.
- Developed word-unit scoring with original-order rendering for budgeted evidence materialization to preserve document structure under compression, increasing F1 by 13.2% on Natural Questions and reducing compression overhead by 96.7% compared to Selective Context.
- Optimized the online path of score lookup, quota allocation, and budget repair for interactive RAG serving, reducing compression latency by 98.9% compared to EXIT and 99.7% compared to FaviComp.
Multi-agent Research Synthesis Engine
Undergraduate Research Assistant, 2025.11-2026.05, Champaign, IL
- Orchestrated eight specialized agents across a 12-step workflow for automated literature synthesis to cover planning, retrieval, drafting, reflection, and safety review in a single loop with LLM-as-judge evaluation, achieving 95.5% accuracy on deep research pipelines.
- Optimized Semantic Scholar and Tavily tool calls for multi-source literature retrieval to reduce latency on the critical path with a bounded thread pool and typed fallbacks, reducing query time by 40.2%.
- Built Model Context Protocol servers for agent tool access to centralize governed tool integration behind one typed interface over academic databases, code repositories, and document stores, reducing integration latency by 87.5% across eight agents.
Links: Project Page