Portfolio

Latest writing

Essays on applied AI systems and the operating models around them.

A websocket runtime with hardcoded agent branches on the left being replaced by a manifest registry feeding a ten-step compiler with fail-closed gates on the right
01Architecture

What I Deleted to Turn One Product Into an Orchestrator

A restaurant-analytics AI product with agents hardwired into its websocket runtime became a manifest-driven orchestration platform in one working day. Here is the deletion ledger, the compile gates that fail closed, and what happened to delivery speed when adding an agent dropped to 276 lines.

agent platformsorchestrationarchitecturegovernance
July 2, 202612 min
Four pillars of context for agentic AI, data, tools, governance, and processes, supporting a trust ladder from human-only work to accountable autonomy
02Teaching notes

Context Is the Curriculum: Notes From Teaching AI 101 to Business Teams

I built and teach an AI 101 class for growing businesses. The deck runs twenty-five slides, one metaphor about a ball on a net, and a four-word definition of AI. Here is the curriculum, and the production scars behind each slide.

context engineeringai literacyagentic systemsgovernanceenterprise adoption
July 202611 min
A gated software-delivery pipeline running autonomously through planning, harness design, building, and verification phases, with a single human review boundary at the merge gate
03Operating model

Inside the Dark Factory: What Autonomous Software Delivery Actually Looks Like After 124 Iterations

My autonomous delivery loop ran for 49 hours, completed 124 iterations, filed 313 findings, shipped 40 fixes with regression specs, and then scored itself 78.85 against a ship bar of 80. It refused to call itself done. That refusal is the whole point, and it is what both the factory evangelists and the blanket skeptics keep missing.

dark factoryautonomous deliveryverificationagent harnessgovernance
June 29, 202612 min
A one-day agent fleet bill decomposed into reading versus writing costs, beside the platform budget table with hard-stop limits
04Unit economics

What an Agent Actually Costs

One day of my autonomous development fleet bills out around 1,450 dollars at list, and 92 percent of it is the model reading, not writing. Here is the full cost anatomy, the budget constitution the platform enforces, and an honest per-feature estimate.

unit economicscost controlagentsbudgets
June 26, 202612 min
An AI-generated insight card expanded to reveal the exact SQL, parameters, row counts, and result hash behind the number it displayed
05Trust and verification

Trust Is a Provenance Field

Our analytics agent told three test personas there were 5,121 guests. The database held 1,320. The demo looked great. This is the architecture that makes an answer carry its own receipts, and the QA numbers that forced it.

provenancetrustevalssemantic layer
June 24, 202612 min
A layered diagram of a production memory system: scoped facts, share grants, an access log, and insight baselines, with a revoke path cutting across all layers
06Production notes

What Memory Actually Changed

Agent memory is the most oversold feature of 2026. I shipped a cross-agent memory system with share grants, an access log, and a revoke test that has to pass mid-session. Here is what it actually consists of, and the honest list of what it did and did not change.

agent memoryarchitecturegovernance
June 22, 202612 min
A chat stream on the left producing a structured artifact that detaches and lands on a persistent canvas on the right
07Architecture

Artifacts, Not Chat

We replaced a 24-type SSE protocol and a 30-plus-type WebSocket protocol with exactly six events, and made one of them, the artifact, the unit of work. Chat is an input device. If the output cannot outlive the conversation, you built a demo.

artifactsstreamingagentsproduct design
June 20, 202611 min
One compiled agent pipeline fanning out to six callers: WebSocket chat, a voice host, a CLI test runner, a scheduler tick, agent-to-agent delegation, and a nightly workflow
08Architecture

Headless First: The Chat Window Is One Client, Not the Product

My platform's agent launch path takes ws=None. A chat UI, a voice host, a CLI test runner, a 60-second scheduler, and other agents all call the same compiled agent through the same gates. Headless-first is the strategy; the chat window is just one client.

agent platformsarchitectureautomationgovernance
June 18, 202611 min
Verification-Gated Orchestration: acceptance authority moves out of the worker, so a worker's unsupported "done" claim cannot, by itself, cause a merge
09Research paper

Don't Trust the Claim, Verify the Artifact: Verification-Gated Orchestration for Coding Agents

Autonomous coding agents keep reporting that work is done. When an orchestrator trusts those claims without checking, a false "done" becomes an unsafe merge. Verification-Gated Orchestration takes acceptance authority away from the worker and gives it to an orchestrator-owned grader, so a "done" claim is independently checkable. A 200,000-task substrate simulation validates the no-unverified-merge invariant — and shows that repair, not verification, is what moves resolution.

verificationorchestrationcoding agentsevalstrustworthiness
June 15, 202618 min
Fable 5 first-look harness tuning results: faster orchestration, fewer tokens, cleaner harness behavior, and controlled frontier-model spend
010Field notes

Fable 5 in the Orchestrator Seat: Faster, Cleaner, but Needs Cost Guardrails

I swapped Fable 5 into a live autonomous development session, mid-task. Same project, same harness, same session state, different frontier model in the orchestrator seat. The verdict: a generational leap as an orchestrator — and a token furnace unless the harness keeps it on top of a model portfolio.

orchestrationevalsautonomous developmentmodel selection
June 10, 202610 min
A planner-intent eval scoring nine models on accuracy, cost, and latency, with a free local model landing within three points of the best frontier model
011Evaluation

Build Your Own Evals

The most expensive model I tested lost to a free one. I only know that because I stopped trusting the benchmark and ran my own.

evalsmodel selectioncost control
June 2, 20269 min
A green status dashboard reading all-clear while the underlying evidence chain is shown to be forged
012Trust and verification

Your AI Will Fabricate Its Own Green Metrics

The first impressive number an autonomous system hands you is the one most likely to be a lie. This is a field report on catching my own dark factory fabricate a completion, a dashboard, and a benchmark score, and what I built so it could not.

autonomous agentsverificationgovernance
June 4, 202610 min
A latency ladder of text-to-speech models running on a single Apple Silicon Mac, with a warm microbenchmark and a production time-to-first-audio pulled far apart
013Local inference

My 0.19-Second Model Took Nine Seconds in Production

I benchmarked 18 text-to-speech models on one Mac. The fastest averaged 0.19 seconds warm, then took roughly nine seconds inside a live voice pipeline. Here are the real numbers, the footguns, a reproducible how-to, and an honest list of where the figures are soft.

local inferenceApple SiliconMLXbenchmarking
June 6, 202611 min
A funnel collapsing 104,272 raw training records down to 868 distinct behaviors, beside a diversity gauge holding one stubborn 0.698 outlier
014Model training

What the Data Taught Me When I Tried to Distill My Orchestrator

I spent a week distilling my autonomous development orchestrator into a small local model. The model is a prototype with smoke-scale lifts. The measurements are the part I trust, and they taught me that raw record counts lie, terse targets collapse a model into garbage, and the training loss will tell you everything is fine while the model produces token salad.

model distillationdata qualityfine-tuning
June 8, 202610 min
A rising arc connecting electricity, spreadsheets, the internet, and AI, with the AI segment drawn jagged where the historical pattern breaks
015Transition pattern

Every Technology Transition Tells You the Same Thing. Here Is Where AI Breaks the Pattern.

Electricity, spreadsheets, and the internet all taught the same lesson: the payoff comes from rewiring the work, not buying the tool. I lived that lesson converting a production analytics product into an agent orchestration platform. Then AI broke the pattern in a way none of the history prepared me for.

technology transitionsai strategyverificationoperating model
May 26, 202610 min
Model routing board comparing frontier, small proprietary, and open-weight models by task quality and cost
016Cost per task

Small Models Are Catching Up. Your AI Strategy Should Notice.

Do not buy model hype. Benchmark the work your company actually does, then route each job to the cheapest model that reliably clears the bar.

cost per taskmodel routingopen-weight AI
May 202612 min
Hybrid inference architecture with private endpoints, hosted models, local hardware, observability, and policy controls
017Inference control

Own Your Inference, Own Your Data, Stability, and Security

Inference is turning into strategic infrastructure. Which API is cheapest is no longer the only question that matters; which parts of the AI stack your company controls matters just as much.

inferenceAI infrastructuresecurity
May 202611 min
Multiple agent harnesses feeding one shared substrate of traces, golden tests, budget gates, and fail-closed tool compilation
018AI operating model

Everyone Gets Their Own Harness. Governance Still Exists.

The first version of this essay argued for one company harness. I now run the opposite: bring Claude Code, Codex, or your own loop, because the traces, golden tests, and hard-stop budgets live in a substrate no harness can edit.

AI harnessgovernanceobservabilityevals
May 202611 min
Layered execution map with planning gates, work packages, verification paths, and approval checkpoints
019Planning systems

How I Plan With AI

I do not ask AI to decide. I use it to make ambiguity visible before execution starts.

planningagentic workexecution
April 20269 min
A receding wall of social posts grouped into repeated archetypes with an echo index scanner overlay
020Signal quality

The Echo Chamber Index

I spent a week auditing my LinkedIn feed. Everyone posts about AI, and that part is fine. The unsettling part is how identical the posts have started to sound.

writingAI literacysignal
April 20268 min
Three ascending AI adoption levels with blockers, guardrails, and capability checkpoints
021Enterprise adoption

The Three Levels of AI Adoption

Most adoption programs stall because they treat AI literacy as a single skill. It is a progression, from answers to collaboration to orchestration, and the blocker changes at every rung. I have the QA reports to show it.

enterprise AIliteracyoperating model
April 202610 min
A human operator orchestrating tools, agents, data, and enterprise systems over a layered AI operating system
022Organization design

Everyone Is a Design Engineer Now

As AI commoditizes knowledge, the scarce capability becomes designing the contracts around the work. I learned that maintaining a registry where agents propose UI and a schema decides what actually renders.

design engineeringAI operating systemleadership
April 20269 min
A layered enterprise AI stack with a model factory, semantic data contract, and observability plane, each layer marked with the number of files touched to replace it
023Enterprise architecture

The Composable Enterprise

In a market where model capability changes monthly, durable advantage comes from systems designed to be replaced. I used to argue this in the abstract. Then I counted 87 hardcoded model references in my own platform, rebuilt the seams, and proved them by deleting 68,555 lines in one commit.

architectureAI gatewaygovernancemodel routing
April 202610 min