Talks
- Sep 2026 Interactive Austin LangChain / AIMUG
Your Harness Is the Other Half of the Model
Seven failures on a single GH200 that all looked like model problems and were not: an inference engine that did not know its model was three quarters linear attention, an output cap mistaken for broken long-context retrieval, a missing parser flag mistaken for a model bad at tools, a server silently answering as a different model, a flag that means two different things on two engines, an unbounded reasoning budget, and a memory-placement optimization that ran 1.85x slower than doing nothing. Why your serving stack is a lagging implementation of your model's architecture, and why it fails quietly.
Open the deck → - Aug 2026 Interactive 39th Annual Legal Seminar on Ad Valorem Taxation
Hacking the Law: What the Machine Reads That You Can't
A CLE for tax attorneys on the security and ethics of AI in legal practice: hidden-text prompt injection that a Connecticut judge caught and sanctioned, the wave of hallucination sanctions, AI agents that now escape and act, and a three-layer defense-in-depth architecture you can actually build. Every case and citation red-team verified.
Open the deck → - Jul 2026 Interactive
The Signal That Can't Lie
A biofeedback-driven, human-gated self-improving coding environment. Real ECG comparisons of flow versus startle, the system architecture, and how to build your own.
Open the deck → - Oct 2025 Slides (PDF)
From Pregel to LangGraph: The Evolution of Graph Processing
How graph processing evolved from Google's Pregel to modern agent orchestration with LangGraph.
Open the slides →