I work on ML systems, graph models, and interpretability.
Current focus: objective monitoring under model modification.
A Protocol for Lineage-Invariant Objective Probing - preprint protocol for testing whether objective-predictive internal probes transfer across modified model descendants without mostly learning lineage shortcuts.
Custody - provenance and revocation for agent memory on Google ADK and Vertex AI Memory Bank, deployed on Cloud Run with Firestore. Found and fixed a real bug in ADK's Memory Bank service, google/adk-python #6946 and #6947.
BrainConnect-ASD - adversarial graph neural network for site-deconfounded ASD classification from ABIDE resting-state fMRI.
Python 路 PyTorch 路 Rust 路 Go 路 PostgreSQL 路 SQLite 路 Docker 路 AWS 路 Google Cloud Run


