Hello World

Welcome to my working notebook for Oxygen AI.

This site is where I am going to write down the parts of AI engineering that deserve more careful language: memory, intent, agency, measurement, safety, and the strange physics of systems that learn from interaction.

I have spent years building software close to real constraints: production systems, interfaces that need to be trusted, and AI infrastructure that has to behave when the demo is over. The deeper I go, the more convinced I become that the next useful frontier is not just bigger models. It is better memory, better measurement, and better control.

Why Oxygen AI

Oxygen is a useful metaphor because it is essential, quiet, and dangerous when mishandled. AI should become infrastructure in that same sense: present enough to help, disciplined enough to trust, and designed with safety as a first principle.

I am interested in systems that can remember without becoming brittle, adapt without drifting away from their purpose, and learn intent without requiring endless prompting.

That is the thread tying this journal together.

What I Will Write About

Some posts will be practical. Some will be speculative. Some will be stories I can only tell in outline for now.

Memory systems that give language models continuity without handing them uncontrolled autonomy.
Intent learning and the question of how an agent knows what a person actually wants.
Drift and identity vectors as measurable signals for safer alignment.
Tiny agents that converge quickly because the measurement loop is right.
Production systems where correctness, observability, and reliability are not optional.
AI interfaces that make uncertainty visible instead of hiding it behind polish.

My Bias

I believe the best AI systems will be judged less by how fluent they sound and more by how well they preserve context, recover from ambiguity, and stay inside their intended identity over time.

That means we need better engineering around memory. We need evaluation that can see drift before it becomes failure. We need agents that can update their behavior without losing their safety envelope.

This journal is a place to build those ideas in public.

Join the newsletter below, or follow along as the notes become more technical.

Let's build systems that can breathe without burning the room down.