Repeatable, verifiable AI-assisted software engineering. I build software with AI agents under engineering discipline — specs before code, gates that can refuse, review layers that catch what the previous layer missed — and I publish the receipts so the claims can be checked rather than taken on faith. Every repo below runs its own tests in CI; the numbers in each README are the numbers the badge proves.
- Verification, not trust. A completion claim carries evidence or it isn't a completion claim — commit SHA if code changed, test command and exit code if tests ran. Schemas enforce it, because prose drifts, schemas don't.
- Claims corroborated by commits. Execution reports are assembled from the run journal and
git log; nothing is asserted without a commit to point at. - Test against your own confirmation bias. Hand-authored expectations are written by the same person who wrote the code, so they get checked by differential tests against independent implementations and property-based invariants. That layering surfaces real bugs.
- Gates that can refuse. Plan validation returns FAIL and execution stops; only an explicit human override proceeds, and the override goes in the report. AI is a force multiplier for engineering judgment, not a substitute for it.
- Regenerate over patch. When a codebase is bad enough, capture its behavior in tests, then rebuild clean-room from that specification instead of patching around the defects.
Verifiable delivery · opening-dojo — a chess opening trainer built end-to-end by a multi-agent pipeline, with the PRD, plan, parallel agent lanes, five review passes, traceability matrix, and postmortem all committed. 86 tests in CI.
Regenerate-don't-patch · gorules → rustrules — a 2016 Go rules engine audited by four AI specialists (37 findings), test-rehabbed 29→124 to pin down real behavior, then regenerated clean-room in Rust: 12+ bugs fixed, 15 operators added. 124 and 121 tests in CI.
Applications · the-strand — a Go literary-analysis engine with a provenance-linked pipeline, deterministic replay providers, and release gates · tonewise — a native iOS ear trainer teaching relative pitch numerically, 93 tests over the domain engines.
More publishing as it clears review — MCP servers, evaluation harnesses, and the methodology skill library behind the workflow above.
Work runs on a five-machine homelab: a Ryzen 9950X3D / RTX 5090 workstation for local inference and agent runs, an Apple Silicon node for always-on inference, and an Ubuntu server for reverse proxy, monitoring, and storage — with local LLM serving, vector search, and speech-to-text self-hosted.
📍 Boone, NC · savviety.ai