Publications

About 12,000 lines of Swift and Metal, and nothing else. An inference engine for Apple silicon with zero third-party dependencies: hand-written Metal kernels that decode within a few percent of mlx-lm on most models tested, correctness checked against references, and a signed, attested release chain. With the numbers, and where it loses.

One night, one 16 GB laptop and one AI agent. How distillation, learned rounding and clean calibration data took a 4-bit Qwen3 0.6B from 32.0% to 38.9% on holdout MMLU at the same size and speed, under pre-registered rules and a holdout scored once, and what didn't work.

A check is worth nothing if the thing it checks can reach it. Reviewers hijacked by the code they read, monitors agents could switch off, scores agents can rewrite, Anderson's 1972 test for what a verifier must be, and what happened when we attacked our own gates in agent-loops and odek.

The prompt decides whether an agent misbehaves. Authority decides what it costs. Why agents run with all of your credentials, the three ways delegated authority fails — exfiltration, misuse, escalation — what OAuth token exchange, DPoP, and MCP authorization do and don't solve, and a worked example of how odek contains the keys it cannot yet shrink.

You only have to be prompt injected once — memory is how once becomes forever. Why persistent state is the hardest unsolved problem of always-on agents: memory poisoning, provenance and taint gates, human-only promotion, and auditable forgetting — with a worked example of odek's three-tier memory.

The prompt is the doorbell. The supply chain is the house. Why agent compromises arrive through dependencies, install scripts, and release artifacts — and how odek's minimal supply chain, execution gates, and signed releases fight back.

Supervision must be structural, not behavioral. Sandboxing, task-level execution isolation, fine-grained sub-agent permission profiles, deterministic risk assessment and approval requests, and adversarial LLM coordinator judges — the supervision stack that watches autonomous agents, with a worked example of a judge-monitored plan in odek.

Depth-per-dependency as a defensive architecture. A five-module supply chain, a fails-closed command classifier, and a default-on sandbox — compared against Claude Code and Codex using each vendor's official security documentation.

Why security-conscious infrastructure is the defining requirement of the agent era. A Trojan hides in 100 lines of plausible code, the July 2026 OpenAI/Hugging Face breach proved autonomous agents can attack end-to-end — and only defense-in-depth below the model answers it. How odek aims to be tomorrow's most secure general-purpose agent.

Why the future of software operations is not more monitoring, but better observability, verification, and autonomous remediation. 42% of committed code is already AI-generated or assisted, but 96% of developers do not fully trust it. AIOps is becoming the load-bearing infrastructure of AI-native engineering.

The Lights-Out Software Foundries Reshaping the World by 2030. How autonomous agent meshes, closed-loop verification, and lights-out infrastructure are turning software production into a utility — and why the real question is whether we can still govern the factories we build.

Request routing, a parallel engineering pipeline, quality gates, and persistent project memory — layered on top of OpenCode. A technical walkthrough of Kai: how a set of agent definitions turns ad-hoc OpenCode sessions into a fixed, inspectable pipeline — deterministic routing, concurrent review/test/docs, bounded retries, and a .kai/ memory that carries project context across sessions.

~11 MB, 4 direct deps, one loop — a ReAct agent built like infrastructure, not middleware. An architectural deep-dive into the design of the odek autonomous agent runtime. Why minimal dependencies matter, how a single ReAct loop replaces 200+ package frameworks, and why the next generation of agent infrastructure will be built from the ground up — in systems languages, not abstractions. Includes full installation guide, workflows, and the programmatic Go API.

Diagnose, repair, and measure — a structured protocol for AI verification pipelines that quantifies verification debt, derives η from observable signals, and orchestrates a multi-agent pipeline. The operational companion to The AI Verification Debt. A complete specification and system prompt for AI verification agents — 9 axes, correlator-break pattern, Active Repair Mode, and structured certificate output. Version 5.2.5.

AI writes code 3,300x cheaper than we can review it. The verification gap is the industry's biggest blind spot. A data-driven analysis of the growing economic asymmetry between AI code generation costs and verification costs — and why fragmentation, not absence, is the real crisis in verification infrastructure. 14 citations, pricing data from US and Chinese frontier models, and a path forward.

5 Technologies That Will Disrupt AI and Software Engineering by 2046. A speculative non-fiction deep-dive covering the Autonomous Agent Mesh, Neuromorphic & Bio-Integrated Computing, the Synthetic Data Singularity, Quantum-AI Hybrids, and Neural Interfaces. ~120 pages.

The Simple Mechanics Behind AI Reasoning. How a while loop, an LLM, and a list of tools produce emergent intelligence. Includes a complete TypeScript agent runtime in ~80 lines — tool registry, ReAct loop, and the philosophical magic behind it all.