I make complicated systems tell the truth about how they work.

§ 01  —  Selected proof

The tools change. The work keeps its shape.

• Factory floor → local AI Factory floor to local AI job shop Factory scheduling in 2019, GPU and agent scheduling in 2026. Different machines; the same discipline of making scarcity and constraints visible.
• Capacity claim → causal result The hardware wall that was software A capacity conclusion retracted, one dispatch constant identified, and the ninth-request cliff moved experimentally.
• Small world → living system Software nobody assigned Community systems that grew from small-world tooling into software supporting living environments and the people stewarding them.

§ 02  —  How I work

Five things, in order. Nothing else.

1. 01 Land Take the system that works on a laptop and make it work where your customers are.
2. 02 Enable Sit next to the engineers who will keep it running, until they would rather I left.
3. 03 Support Hold the pager through the first weeks of real traffic, and write down what breaks.
4. 04 Document Leave behind plain prose a new hire can read on a Monday and understand by lunch.
5. 05 Leave Step back cleanly once the people who own it no longer need me.

§ 03  —  What fits

The work I take.

Almost everything I do is the unglamorous middle of an AI project — after the demo lands and before the team feels comfortable owning it.

• A model or pipeline that works in a notebook and has never survived a real week of traffic.
• A retrieval system whose answers have started to drift and nobody can say exactly when or why.
• An evaluation harness that everyone agrees should exist and nobody has had the time to build.
• A capable engineering team that has not shipped a machine-learning system before and would rather not learn alone.
• A handoff from a vendor, a contractor, or an employee who is leaving, where the documentation is thin.

What I don’t take

• Greenfield strategy, roadmaps, or decks for the board.
• Building a model from scratch when an existing one would do.
• Work whose success metric is the word “innovative.”
• Engagements longer than three months without a renewal conversation first.
• Anything that requires me to pretend a problem is simpler than it is.

§ 04  —  How I work, in practice

Six things I’ve written down, in case they’re useful before we talk.

• What Consumer Hardware Actually Does Benchmarks · Measured Twenty-nine indexed local-model labs plus the August rebuild and Qwen 3.8 campaigns on 12 GB NVIDIA cards and a pair of Intel Arc Pro B70s. The rebuild separated three walls: placement at load, Vulkan dispatch under concurrency, and KV plus attention at depth. Every number dated and attributed — including the corrected retractions.
• Build Your Own HEARTH Tutorial · Hands-On A local MCP gateway for Claude Code on a Windows machine with an NVIDIA GPU. Every tool call logged to a queryable ledger, a local model doing the grunt work, and a test gate at every step. ~90 minutes from nothing to receipts.
• threat-trace Interactive · In Browser A working multi-agent log-analysis pipeline with full provenance chains. Three streams under attack, three discrete agents, one cross-stream layer that catches the false-positive trap a single-prompt scanner walks into. Open it in your browser. No install.
• The Mech Suit Methodology Essay From copy-paste to multi-agent orchestration. How AI-paired development actually works in practice — the methodology underneath everything else.
• Find Every Reason to Dismiss Me Adversarial Review I commissioned an adversarial AI review of my own code, instructed to find every defensible reason to dismiss it. Why I published the result, unedited.
• The Tight Loop Field Report A field test for the Mech Suit Methodology. A 3D replay viewer built in two slices over two days, with the receipts.

More writing →

§ 05  —  Terms

The discipline of a small shop, written down.

• One client at a time. If I am working with you, I am not working with anyone else.
• Fixed fee, paid in thirds: at the start, at the midpoint, and on the day I am done.
• Three months is the longest engagement I will quote without an explicit renewal.
• One written update each week. No standing status meetings, on either side.
• I work from Pacific Time, mostly mornings, and answer email within a business day.
• I will sign whatever non-disclosure agreement your counsel wants. I do not sign non-competes.
• If the work turns out to be wrong for me, I will tell you in the first two weeks and refund the difference.