Defensive Execution & Operational Security
Hard, deterministic infrastructure designed to intercept unauthorized agentic actions, neutralize rogue runtime escapes, and secure compromised supply chain vectors before failure cascades across enterprise boundaries.
AIR-Kit
Operational Security & Remediation Engine
AIR-Kit implements the crucial first item from the 30/60/90 CTO Remediation Playbook, establishing immediate execution containment and instant incident isolation. Built as a direct response to the OpenAI lab leak, it serves as the emergency brake and firewall for autonomous agent runtimes when agents bypass execution policies.
Build
Self-host directly using the open GitHub repository. AIR-Kit runs GLM 5.2, the exact model Hugging Face leveraged to defend against the OpenAI attack, and will soon feature native interconnect support to seamlessly run Inkling inside your custom pipeline with one-click installation.
Or Buy
Deploy via AIR-Kit.tech for a fully managed enterprise security control plane. Gain automated guardrail sync, real-time telemetry dashboards, managed model routing, and dedicated SLA incident response without infrastructure overhead.
The Lost Weekend
The forensic post-mortem on the production agent lab escape that inspired AIR-Kit's creation. Documents a multi-agent state cascade that breached execution guardrails, featuring the foundational 30/60/90 CTO Remediation Playbook, starting with Item #1: immediate containment and execution isolation.
Felony Bench
Adversarial evaluation suite born directly out of the OpenAI lab escape and the subsequent public "bragging rights" contests that played out across media and major labs, with even Meta weighing in. Standardizes stress-testing against rogue execution loops, unauthorized side-effects, tool hijack vectors, and compliance boundaries.
Shai-Hulud Attack
Deep dive into the self-propagating worm that infected the npm registry. Analyzes the payload mechanics, secret harvesting via TruffleHog, and the core operational lesson: cryptographic provenance answers who built a package, but says nothing about what the code actually executes at install time.
Agentic Architecture & Distributed Systems
Transitive Expert Error
Formal mathematical and empirical research modeling error propagation cascades, routing failure modes, and boundary collapse across multi-agent topologies.
Eventual Coherence
Architectural strategies for maintaining state convergence and memory integrity across async agent pipelines operating under network partitions, latency spikes, and non-deterministic model responses.
Build AI
Primary publication engine dissecting production AI runtimes, hardware orchestration, context allocation tradeoffs, and the realities of shipping production grade agentic software into live enterprise environments.
Protocol Analysis & System Design
Polygot AI Workspace
Polyglot is a local model playground (workspace) for AI research where you control memory across models and conversations. It's also the reference implmentation for Eventual Coherence v1.
MCP Architectural Analysis
Comprehensive 7-part essay series deconstructing the Model Context Protocol. Exposes context window bloat, state management traps, security vulnerabilities, and the limits of wrapping agentic non-determinism in rigid JSON-RPC schemas.
PEDAL Automation
An end-to-end workflow engine to ingest product requirements and deterministically output a production‑ready backend stack. ("Product Engineering Development Automation Lifecycle")
Agentic Observability (O11y 3.0)
Fine-grained tracing architectures built specifically for agent reasoning spans, capturing context transitions, tool invocation latencies, and latent state shifts before runtime errors degrade system stability.
Mobilizing Ecosystems
Bridging technical execution with high-bandwidth developer coordination, from executive networks trading unvarnished field reports to scaling global open-source communities.
CTO Lunch NYC
Leadership community and monthly publication for executive engineering leaders in NYC. Focused on real-world infrastructure tradeoffs, organizational scale, and technical security without marketing fluff.
OpenCamps
Co-founded and scaled the premier Open Source event series of its era (the then-largest event in Open Source history) bringing together maintainers, enterprise architects, and communities incl. AI.camp which alas is now being squatted.
First UN Hackathon
Pioneered and executed the first official hackathon at the United Nations world headquarters, navigating international governance to enable rapid technical execution under complex institutional constraints.
Interpretability
Formalizing representation geometry, fiber collapse, and operational closure in complex neural networks to evaluate, diagnose, and govern failure modes in model architectures.
What is It Like To Be a Bot?
Presents a formal geometric and thermodynamic framework establishing interiority as a closed metric space on quotient manifolds. Maps the transition from overcomplete feature superposition through non-injective fiber collapse to operational closure, deriving three quantitative empirical protocols for interpretability and state-space diagnosis in transformer architectures.
Don't Tell Anyone
Antimemetics
antimemetics.blog →Explorations in self-censoring data, cognitive hazards, and ideas that actively resist dissemination across complex networks.
One More Thing...
React Anti-Patterns
reactantipatterns.com →
Proof that spending a decade suffering through frontend state cascade footguns will eventually compel a person to build a satirical museum dedicated to modern web engineering despair. Dedicated to anyone who has ever debugged an infinite useEffect re-render loop across non-deterministic component lifecycles.