Research essays, field notes, and case studies on practical AI systems for workflow improvement, operational intelligence, and decision support.
Recent writing across practical AI systems, operations, and decision support.
Every AI coding session has a hidden token drain: large tool output from JSON responses, file reads, and test failures fills the context window with structure the model doesn't need. Measured across real-world sessions, automatic compression via TOON notation and debug-output collapsing recovers 24–66% of that spend.
Read articleFTS5 alone hits 95.40% R@5 on LongMemEval-S. Adding a 384-dim ONNX vector pipeline via weighted fusion dropped it to 82.40%. Here's why, and what it means for how you build memory into AI.
SaaS gave everyone the same tool. AI gives everyone a custom one. Here's what the shift means for how software gets built, sold, and maintained.
Construction is full of documents, revisions, handoffs, and tacit judgment. Practical AI can help teams understand project risk earlier and make better operational decisions.
AI becomes genuinely useful when it helps teams understand workflows, domain expertise, and proprietary operational data well enough to improve execution and support better decisions.
UrbanReg AI shows how practical AI can help teams review urban renewal and building-related documents faster, surface regulatory risk earlier, and support higher-confidence decisions.
The biggest productivity killer in AI coding isn't insufficient AI capability — it's starting from scratch every session. Codex Factory Kit fixes this with durable workflow artifacts.
The Claude Code skill library grew from a handful to hundreds. That's supposed to be a good thing — but it created a new problem: tools you don't remember to use might as well not exist.
AI tools doing dangerous things isn't the new problem. The real question is whether you'll know before it happens. WardnMesh sits between the AI and the operating system.
An AI that can't prove what it did is not ready for the real world. Audit trails are not a feature — they're a prerequisite.
When agents hire agents, traditional authentication breaks down. My understanding: trust must be earned per transaction, not granted per credential.
Prompting a model to 'be safe' is not governance. Governance is what happens when the model misbehaves anyway.
Most AI tooling ships with dozens of dependencies. MeMesh ships with three. This is a deliberate philosophy, not an oversight.
AI that assumes a 'fresh start' user inevitably breaks down outside demos. Usable AI must work with accumulated experience.
Building a business shouldn't require a CS degree. A manifesto for the non-technical founder.
Transforming Claude Code from a stateless assistant into a project-aware AI teammate.