Agentic Knowledge Graphs for Research Progression — Orchestrated agent system (Navigator / Extractor / Continuation) building an LLM-curated research KG --- problems, assumptions, constraints, evidence spans, datasets, metrics --- as the provenance substrate for downstream research-assistant responses. Denario agent framework; Python / FastAPI; staging deployed on Cloud Run. Approaching experimental comparison against LLM-generated KGs.
VTTSI — LLM-Adaptive Intersection Safety Scoring — Distributed rewrite of a CS6604 traffic-safety application: provenance-tracked evidence ingestion (NWS weather, VDOT/511 cameras) → deterministic RT-SI + MCDM baselines → Claude Opus adaptive adjustment with grounding guardrails and deterministic fallback → honest-null ablation eval (Spearman / Pearson). Python ports-and-adapters microservices; live-Claude integration wired; ~97% offline coverage.
Construction-AI: LLM- and KG-backed Material Takeoff — Web app parsing architectural drawings (DWG / DXF / PDF), extracting geometry via PyTorch / OpenCV / EasyOCR, producing JSON material takeoffs with OR-Tools cut-list optimization backed by a Neo4j LLM-curated KG. Phase 1 MVP deployed on Cloud Run; Phase 5 in progress adds a VVUQ (verification / validation / uncertainty quantification) layer for engineering-simulation reproducibility.
