Build log
The research isn't hypothetical — it comes out of these. Systems in active development, tools in daily use, and published work.
A local-first, voice-first AI resident: one Python process on one box, with a DuckDB graph for memory. You talk to him, and you can interrupt him. He reads open sources, notices the gaps in what he knows, checks his claims against a council of models, and changes nothing in the household without a PIN, a backup and an audit trail.
The case for an AI resident →
Medical Agentic Resident Assistant (MARA): MAX pointed at healthcare. It carries the record, coverage and history across everyone who touches a patient, hears the answers in the triage call, flags the question nobody asked and the claim that will not hold. Like a resident it prepares, drafts and checks; a person makes every clinical, coverage and payment decision.
Launch the interactive demo →
A dependency-free scaffold that stands up a whole software engineering organization run by LLM agents, from architecture to release, as one written and auditable process. Status is counted from the repository, CI and an append-only ledger, never asserted by an agent about its own work. The point is to find the decisions that have to stay human.
Open the interactive demo →
Local-first network visibility and gateway defence for someone never trained to read a packet. A fleet of sensors reports to one fusion layer that turns corroborating signals into a single plain-language card, then proposes the fix and acts only on your approval: every change PIN-gated, backed up and reversible.
Read the field note — the porch camera’s heartbeat →
FeedHackerA free Chrome extension that clears the noise from your LinkedIn feed. A Mute / Solo mixer hides AI slop, ads and engagement bait, and an on-device scorer explains every flag and learns from your corrections. Nothing is deleted and nothing leaves the browser: one permission, no account, no tracking.
Add to Chrome — free →
Ten pieces that teach a technology leader the language finance already speaks: capital versus operating cost, capitalising the work on a Jira board, depreciation, placement, vendor renewals, AI spend, headcount, and outsourcing after Section 174A. Each stands on its own; read in order they make one argument.
Start with part one →A 27-part series that maps production-grade agentic AI end to end, written from inside the build: the nine-layer stack, MCP and A2A, the RAG family, memory and context engineering, and multi-agent patterns. Each part carries its diagram, its trade-off and a decision framework, including when not to reach for an agent.
Start with part one →The site you are reading: plain HTML, CSS and JavaScript with no framework, published by a standard-library Python builder from one catalog and the source Word documents. Its search runs in your browser, and its research assistant answers only by quoting the articles, so it cannot invent an answer.
An animated cognitive map for an AI dashboard: nebula clouds for modules and synaptic lines for the connections between them, drifting on an HTML canvas. It ships as a drop-in ES module with a small API, and shows a still frame to anyone who prefers reduced motion.
A colony simulation in the Settlers and Warcraft lineage where the workers are hamsters, and there is no game over. Run a real economy across seven biomes with seasons and weather, raise families of named rodents, and recover from raids and floods. Vanilla JavaScript on one canvas, every sprite generated at runtime.
Play the preview — 1:04 →
Nothing in this category yet — check back after the next push.
New research, projects and field notes as they publish. Following happens on LinkedIn — we never see an address, and there is nothing to unsubscribe from.
An occasional summary of new research, each with a link to the full piece.
Confirmed opt-in: we send one link to check the address is yours, and you are not on the list until you open it. Every email carries a one-click unsubscribe that removes you from the newsletter only — never from LinkedIn.