I architect local-first AI systems and direct implementation agents through specifications, adversarial QA, instrumentation, and benchmark-gated iteration.
I don't hand a prompt to a model and ship what comes back. I run an adversarial engineering loop and hold the authority at every gate:
hypothesis → specification → implementation agents → instrumentation
→ benchmark → falsification → diagnosis → revised architecture
The receipts for that loop are in the issue trackers, not just the code: documented null results, isolated host variance, frozen benchmark versions, and issues I file against my own flagship features when the measurements falsify them.
| Tier | Repo | What it is |
|---|---|---|
| 🚩 Flagship | Cymatix Context | Local-first context/retrieval engine for LLM agents, no LLM call on the retrieval path. Shipped Python package; CLI + HTTP + MCP; ~3,400 tests; reproducible benchmark receipts. |
| 🔧 Operational tooling | ScoreRift | Dual-layer audit: automated scoring + manual grading + reconciliation. The product isn't the score, it's watching the gap between them. |
| 🏗️ Applied-architecture demos | DriftWatch, BookKeeper | Reference architectures — SOC 2 draft-generation assist, and a double-entry ledger core. Architecture samples, not certified/production products. |
| 🧪 AI-development laboratory | BigEd | Explicitly not a product a lab for local-first multi-agent orchestration on ~11 GB VRAM edge hardware, file-based handoffs, HITL gates. |
| 🔱 Forks & benchmark fixtures | headroom, AssetOpsBench, anker-solix, archestra-fork | Upstreams I build on or benchmark against. Labeled as forks, not pinned as my work. |
Human (me): product thesis, constraints, architecture selection, acceptance criteria, experiment design, risk decisions, QA and falsification authority. Models: implementation, refactoring, draft documentation, test generation, code-review lenses.
I don't claim code authorship I don't have, and I don't donate the architecture authorship I do.