GraphRAG Hybrid RAG
A retrieval system that knows which kind of question it's answering.
Routes queries between vector search and graph traversal depending on question type, backed by a full evaluation harness that compares retrieval strategies head to head rather than assuming one wins.
View repositoryPortfolio Health Reporting Agent
An agent that reads messy spreadsheets and tells you what actually needs attention.
A data-resilient AI agent for project health reporting: multi-schema Excel parsing, a five-signal weighted RAG framework with manual risk overrides, and automated monthly executive summaries.
View repositoryGTM-Flow-AI
Turning a company website into a structured fintech-alignment score, automatically.
An AI data pipeline that scrapes corporate websites, extracts fintech-alignment metrics using Llama 3.1 via Groq, and structures the results into a local SQLite analytics database for downstream use.
View repositoryPhysio-Topological Dynamics
Fusing brain, heart, and skin signals into one shared shape.
An end-to-end framework for combining disparate bio-signals into unified topological manifolds, with automated task-detection built on top of the fused signal space.
View repositoryUncertainty-Aware Task-Conditioned Point Cloud Upsampling
Ongoing research — unpublished.
A pipeline that lifts 2D affordance and uncertainty heatmaps into 3D, then densifies sparse point clouds conditioned on both the task ("pick up" vs. "pour from") and per-point uncertainty, with a boundary-rejection module to stop points leaking across object edges. Early full-training results show Uncertainty Agreement improving from 0.23 to 0.54 and Chamfer Distance dropping from 0.0138 to 0.0107 over the initial baseline — promising, but still early and unpublished.
Topological Analysis of EEG Phase Synchronization Under Cognitive Load
Ongoing research — unpublished.
Applied persistent homology to EEG phase-synchronization networks across 48 subjects under rest vs. cognitive-load conditions. Cognitive load drove roughly 7.4× more reorganization in global network structure (H0) than in cyclic subnetwork structure (H1), an effect present in all 48 subjects (Wilcoxon p = 7.1×10⁻¹⁵, Cohen's d = 3.48) and robust to a 10,000-iteration permutation test. A two-feature classifier built purely from these topological distances separated rest from task with 95.8% leave-one-out accuracy — evidence that topology captures reorganization patterns that connectivity-density metrics alone can miss.