Side Projects

Projects

One shipped accessibility tool in weekly use — and a multi-month arc of AI work converging on a single question, what does it take to build an AI collaborator you can actually trust?

The homelab is the substrate. What runs on it falls into two groups.

big-ads stands on its own: a shipped product, in weekly use by the one person it was built for, with no AI in it at all — and that was the right call.

The rest are chapters of one story. I kept building AI tools to scratch specific itches — a way to give an assistant current knowledge, a way to give it safe access to my systems, a way to make it a genuine collaborator — before I understood they were three faces of a single problem: what does it take to build an AI collaborator you can actually trust? Cortexa is where that question finally has a name, and it’s the successor the earlier attempts are folding into.

Accessibility · Shipped

big-ads

A grocery flyer viewer for a user with severe visual impairment. Four hostile retailer sites scraped on a schedule, served through a 347-line dependency-free UI. Full GitOps delivery into the cluster.

Knowledge · First-generation

Context Engine

Local-first RAG over files, sites, and git repos — embeddings on my own GPU, vectors in Postgres. Built and CI-tested as a standalone service; its retrieval capability is folding into Cortexa as the knowledge layer.

Governance · First-generation

Policyclaw

A policy gateway that put every mutating AI action behind a human confirmation gate — Go, an OPA sidecar, per-backend risk tiers. A working prototype whose real product was the lesson, now a design input to Cortexa.

AI Systems · The synthesis

Cortexa

An AI collaboration engine that puts collaborative behavior in code instead of a system prompt — plus a blind-judged evaluation harness built to falsify that claim. The first run said it had. The successor the other two are converging into.

---

The through-line

Four efforts, one question. Each solved a piece of “a collaborator you can trust,” as a standalone thing, before I saw they were the same problem:

• Knowledge — Context Engine. An assistant is only as good as what it can actually see. Local-first retrieval answers from current documents instead of a model’s stale recollection. This is Cortexa’s knowledge layer.
• Agency — the platform’s MCP layer. Eleven scoped, credential-contained servers give an assistant structured access to the real systems. These are Cortexa’s hands — and unlike the other two, they stay live as platform infrastructure regardless.
• Governance — Policyclaw. Agency without a leash is a liability. Policyclaw made every mutating action pass a human confirmation gate. The tell that convinced me these were one project: Cortexa’s code-owned memory handshake is the same idea — approval enforced by code, not model goodwill — rediscovered at the collaboration layer.
• The mind — Cortexa. The frame that holds all three: something that knows, can act, and is governed, wrapped in a collaboration loop that’s measured rather than asserted.

Two honesty notes, because the point of this site is that I don’t skip them. None of this is integrated yet — Cortexa today consumes only Postgres and Ollama; the convergence is design intent and roadmap, not shipped software. And “successor” doesn’t mean the predecessors were wasted: each was a working first-generation build whose real output was the lesson. Building them separately is how I learned they belonged together.

And then there’s big-ads — no AI, no research question, just a thing someone needed that now works every week. It’s here to keep the rest honest: I converge on hard problems and I finish and ship.