Technical Director & Applied AI engineer. I run agentic AI delivery for government-scale programmes by day, and build local-first agent infrastructure in the open — runtimes and SDKs where models are replaceable components and privacy is enforced by mechanism, not convention.

Currently building geolytic.ai.

Local-first runtime for natural, interruptible voice conversations — models are replaceable components, the runtime is the product. Proven end-to-end on Apple Silicon with local neural TTS.

Dependency-free SDK for governed Goal → Hypothesis → Experiment → Evidence → Decision → Principle reasoning loops — immutable audit events, human-confirmation gates, and a normative conformance suite with an SHA256-pinned compatibility baseline.

Local-first team activity ledger with member-reviewed sharing — raw activity never leaves the machine, and connectors are privacy fail-closed (direct messages rejected at the adapter). On PyPI as teammates and memberkit.

Headless autonomous web-app builder: drop a BRD-level requirement — a PRD file or a chat message — and it designs, builds, verifies, security-probes, repairs, and deploys a working multi-service app unattended. LLM brain, deterministic hands — control flow is code, and a deterministic gate sits after every phase.

One systems-level runtime (voice, Rust), one protocol/governance SDK (reasoning, TypeScript), two applied agent products (team memory + autonomous builder, Python) — four angles on the same conviction: agent systems should run locally, audit immutably, and ask a human before anything consequential.