Now
What I'm currently working on, learning, and thinking about.March 30, 2026
Building
Multi-agent orchestration for Orqys
Wiring up the DAG-based workflow engine to support parallel agent execution. Context and planning agents can now run concurrently.
Portfolio OS knowledge graph
Rebuilt the D3 force graph with proper link validation. Dangling references no longer crash the simulation.
dbt pipeline recovery at Absinthe
Fixed cascading pipeline failures across campaign event data. Root cause: schema drift in upstream ingestion affecting millions of rows.
Learning
LLM validation loop patterns
Exploring self-correcting generation loops where the review agent can reject and trigger re-generation with refined constraints.
Apple MLX inference optimizations
INT4 quantization on MLX gives 3x speedup over FP16 with negligible quality loss for rewriting tasks. Unified memory is the key advantage.
MCP protocol deep dive
Model Context Protocol is the right abstraction for agent-tool integration. Much cleaner than custom function calling implementations.
Thinking About
- Whether agent orchestration should be DAG-based or event-driven at the edges
- The right abstraction layer for local-first AI — should models know they're running on-device?
- How to make deployment invisible without hiding complexity that matters