Projects

Applied AI systems for complex workflows.

A selection of production-oriented AI systems for reporting automation, document intelligence, and data workflows, with an emphasis on domain grounding, reviewability, and human-guided iteration.

TLF Reporting Automation Agents

A LangGraph-based clinical reporting system that turns TLF shell specifications into generated SAS code and a reviewable reasoning trace. The workflow progressively loads ADaM metadata, macro knowledge, and reference programs only when needed, then combines generation, quality review, human feedback, and SAS runtime feedback into an iterative code repair loop for large reporting packages.

LangGraphAgentic workflowStateful orchestrationProgressive context loadingExecution feedback loop

Trial Document Intelligence

eTMF document classification system that avoids the brute-force LLM pattern of sending every taxonomy rule plus the full document into one long prompt. The workflow projects noisy trial documents into rule-like semantic descriptions, retrieves a compact Top-K candidate set from tenant-specific rule vectors, and uses an LLM-as-a-judge step to make classification cheaper, faster, and more scalable.

Semantic projectionVector retrievalRerankZero-shot adaptationLLM-as-a-judge

Clinical Data Management Agents

Clinical data management AI system that supports study setup through multiple focused agents and workflows. The system combines LLM-based protocol understanding and CRF design support, traditional ML for reusable template mining, SFT and specialized SFT for DVP-to-edit-check generation, and DPO-based preference alignment with expert review.

Agentic workflowsLLM orchestrationSFTDPOClustering