Essays, playbooks, and measured findings on building with AI, running agencies, and how work actually gets done - plus personal stories and the opinions I will defend. Much of it published here first, and all of it open.
Ask what the most valuable system in an agency is and you will hear about the CRM, the project tool, the pipeline. Ask what actually runs the place and the honest answer is a skull. The standards, the client history, the judgment, the memory of what was decided: all of it commutes to the office inside one person. Twenty years of fixes failed to move it. I think I finally understand why, and what moves it instead.
A full novel runs 500k to 900k characters. At cloud TTS pricing of roughly $100 per million characters, narrating your bookshelf is a serious invoice. The interesting part is that the math has quietly changed: open-source voice models now run on a consumer GPU, clone a voice from a short sample with zero training, and cost only electricity. So I built the proof.
Oracle shipped a second brain that ingests, embeds, classifies, synthesizes a wiki, and consolidates four kinds of memory on a daily loop. It promises your learned facts never go stale. It delivers on that promise, for one definition of stale: out of sync with a source. There is a second definition, and it is the one that decides whether a system can be trusted over years.
"Use AI where the work is, not where the trust is." The neat version says AI does the labour and you do the real thinking. That is not how it works for me. AI is in my thinking constantly. The line is real, but it is not between tasks - it is between roles.
In early April I published a paper with an unfashionable claim: the AI industry is solving the wrong memory problem. Everyone was optimizing recall. The real challenge is continuity, systems that stay the same expert over time, use current truths, and respect past decisions. Within one quarter, a solo consultant in Mérida and Oracle's engineering org reached for the same vocabulary without talking to each other. That is a problem becoming visible.
AI has been writing most of the code I ship for a while now. Not small snippets - most of it, across multiple applications with real users. When people hear that, they assume I became a faster coder. That is not what happened. I became a different kind of worker, and the question of what is left for the human turned out to have a better answer than I expected.
For persistent expert systems, memory is not just a recall layer. It is the substrate of continuity and trust. Once an AI is expected to persist as an expert across sessions, projects, and accepted prior work, the real failures are no longer forgetting - they are identity drift, stale truths, and contradictions that never die. My paper argues these systems need a different memory architecture entirely.
I write regularly on LinkedIn about what I'm building and learning: agency growth, AI development, product judgment, and the messy reality behind making things work.
Follow on LinkedIn