Research essays and executable reference implementations for applied AI, systems architecture, and production data platforms.
Essays → harveygill.substack.com
Code → here.
Why world models matter — and why token-based AI fails when decisions must persist over time.
The core argument: modern AI failures are almost never the result of bad models. They are almost always architectural mismatches. The alternative: systems built around explicit world models with persistent state, defined dynamics, and verifiable transitions.
Essays: harveygill.substack.com/p/beyond-tokens
Code: MaverickHQ/beyond-tokens
How to build AI systems that can act reliably — deployed, evaluated, and capable of improving from their own experiments.
Built layer by layer: runtime → evaluation → environments → architecture → learning. Each essay mirrors a versioned codebase milestone.
Essays: harveygill.substack.com/p/executable-world-models
Code: MaverickHQ/executable-world-models
The architecture assembled across the series: tokens → models → agents → constraints → artifacts → evaluation → experiments → environments → evidence → policy
Harvey Gill — architect, developer, applied AI practitioner, technical writer.
Large-scale systems design, production data platforms, and applied AI in operational environments. The work here is pragmatic: how systems are actually built and operated, where theory breaks down under real constraints, and what it takes to make AI trustworthy enough to act.
Substack: harveygill.substack.com
LinkedIn: linkedin.com/in/gillharvey
Ideas first. Executable artefacts that make them concrete.