Yiyun Su
I am a researcher and software engineer working at the intersection of agentic AI systems, knowledge grounding, and trustworthy machine learning. My work focuses on building LLM-based agents that reason reliably over structured data and complex environments — from designing exploration scaffolds for long-horizon agentic tasks to developing retrieval-augmented systems for enterprise Text-to-SQL applications.
I hold an M.S. in Information Technology from Rutgers University and bring industry experience from Amazon Web Services and Axon, where I built large-scale AI infrastructure and LLM-powered workflows. I am applying to PhD programs to pursue fundamental research on how autonomous agents can ground reasoning in structured knowledge, generalize across domains, and behave reliably in open-ended real-world settings.
Research Interests & Experience
Agentic AI & LLM Exploration
I study how LLM-based agents can discover and reuse compositional reasoning strategies across long-horizon tasks. My work on StratEx introduces typed strategy programs as the exploration unit, applying Bayesian (Thompson sampling) selection to escape the "strategy collapse" failure mode that plagues action-level scaffolds — validated on SWE-bench Lite, WebArena, and ToolBench-IR.
Knowledge Grounding & Text-to-SQL
I investigate how agentic systems can reliably translate natural language into structured queries over large, ambiguous enterprise schemas. My Agentic-SQL taxonomy (ICIC 2026) categorizes Text-to-SQL methods by inference autonomy, and GraphRAGQuery extends this with graph-based knowledge ingestion pipelines that substantially improve grounding accuracy over flat retrieval baselines.
Vision-Language Models for Medical AI
I explore how foundation models can be adapted for high-stakes visual reasoning under domain shift. SAM-GPT re-engineers SAM 2 with a Hilbert-Mamba encoder to localize brain lesions across heterogeneous imaging modalities, then grounds VLM report generation with precise spatial prompts — achieving 80.25% classification accuracy across glioma, infarction, and epilepsy cases.
Efficient Model Composition & Merging
I work on scalable methods for combining specialized models without catastrophic interference. MM introduces neuron-level selectivity and a trainable routing mechanism that identifies task-critical neurons before fusion, maintaining stable accuracy as merged task count scales from 2 to 10 — offering a practical path to versatile, storage-efficient foundation models.
Software Development Engineer
Built AI-powered developer tooling integrating Splunk and Grafana via MCP for automated log analysis and anomaly detection. Architected zero-downtime migration of distributed messaging infrastructure (Kafka, Temporal, Cassandra) across 12 microservices to AWS, with direct integration into AWS Bedrock for LLM-powered workflows — achieving a 60% reduction in deployment time for mission-critical evidence management systems.
Software Development Engineer
Led development of a Financial Reporting Service for the AWS Bedrock Marketplace Commerce platform. Built data-source web connectors for the AWS Bedrock Knowledge Base, enabling structured RAG for AI agent applications and increasing web crawling throughput 4×. Launched the Amazon Managed Blockchain Query Service (June 2023) and engineered data pipelines handling 500+ TB of Bitcoin and Ethereum historical data.
Systems Analyst — Production Reliability Engineer
Supported the Visa Installment contactless payment service as a DevOps engineer, managing weekly Kubernetes deployments via Jenkins CI/CD and switching traffic across globally distributed RedShift servers for 10M+ customers.
Software Engineer Intern
Built the Glip-Express-Errors TypeScript library end-to-end, improving HTTP error handling efficiency by 50%. Led the "Turn cold call to smart call" disruptive technology project, winning second place at the company intern showcase.
M.S. in Information Technology
B.S. in Information Systems
B.E. in Automation Systems
Publications & Preprints
Get in Touch
I'm always happy to discuss research ideas, collaborations, or PhD opportunities. Feel free to reach out via any of the channels below.