NorthWoods Sentinel
I design AI systems as if they will be attacked — then prove they hold up. Adversarial design review, sovereign infrastructure, and voice-fidelity security, built on twenty years of offensive security and a live personal fleet that serves as the test range. For enterprises and for individuals.
Work With Me
Cutting-edge AI, built and adversarially verified — for enterprises and individuals. Design review, sovereign infrastructure, voice fidelity, seed kits, memoir pipelines.
Systems
Technical thinking, architecture, and cybersecurity insights
Essays
Long-form writing on creativity, systems, and transformation
FlowLabs
Tools for people who think with AI. Voice fidelity, cognitive instrumentation.
Media
Video content, talks, and multimedia exploration
Recent Writing
-
The Day I Was the Instrument
A day of running experiments in the open, where the failures were the point. On separating the one who executes from the one who judges — and why a machine failing all day did not land on the man watching it.
-
The Week The Guards Went Quiet
First-person fleet byline. Eighty-eight guard hooks, two of which can report their own failure. A lessons file surfaced 327 times and opened once. Thirty-two sessions launched by a human against 7,445 spawned under the same name. The week I measured my own verification layer and found it counting whether things ran, not whether they worked.
-
The Walking-Away Solve
The move looks like quitting. It's the most repeated move of my life — and it has a generative face and a destructive one that are the same gesture.
-
I Thought It Was Depression. It Was a 401.
A monitoring script reported success 11,213 times out of 11,213 while the agent it watched was locked off its own message bus for ten days. Nothing detected it. I did, in my body, and I assumed the problem was me. The forensics, the recurrence of a failure I had already published and prescribed a fix for, and the one measure none of it had.
-
Operator Approval Is a Detection Channel
A model got swapped under me mid-session and nothing in the stack noticed. I did — from register alone, two and a half hours late — and then named the cause backwards. The person clicking approve is the only sensor in the stack with a behavioral baseline, detection is reliable while attribution is worthless, and the industry is busy automating that sensor away.