What I'm
Doing Now

Designing and leading the architecture for an LLM-orchestrated agent system handling prior authorization, claims adjudication, and denial management. Multi-agent coordination with specialized clinical, financial, and compliance agents — all running on Anthropic Claude with MCP tool integration.

Model Context Protocol is the missing infrastructure layer for agentic AI. Working through ServiceNow MCP server integration — authentication patterns, tool schema design, context window management for enterprise workflows. The ITSM + AI intersection is underexplored and high-value.

The most fascinating AI research I've encountered this year. Sparse autoencoders applied to Claude Sonnet — extracting millions of interpretable features from superposition. The feature for "Assistant" token activating concepts of restriction and imprisonment is haunting and important. This is the science that needs to precede widespread deployment.

The organ score specifically — there's a version on streaming that strips back the electronic elements and leaves just the pipe organ against silence. I put this on during deep architecture sessions. The low frequencies at 20Hz create a physical sensation that aligns well with the scale of the problems I'm usually thinking about.

When an agent takes an irreversible action — sends an email, modifies a record, initiates a payment — who is responsible? The model, the orchestrator, the human who approved the workflow, the engineer who designed it? I don't think we have good answers yet. Working through a trust architecture framework that treats auditability, reversibility, and escalation as first-class design primitives.

Every new JWST image is an event. Currently following Europa Clipper's journey to Jupiter — it launched October 2024 and won't arrive until April 2030. The possibility of life in a subsurface ocean twice Earth's volume, 628 million km away, is the most important open scientific question of my lifetime.