AI Systems Operator & Automation Architect
Systems that keep their own receipts.
I build local-first infrastructure for AI agents and automation — published on GitHub, documented to the claim tier, and honest about what is proven and what is not. If it is not in the ledger, it did not happen.
AEGIS
A local-first agent operating system for AI coding work — receipt-backed token ledger, an honest-yield rule, a frozen /v1 API. v1.2.0, public on GitHub.
Read about AEGIS →Memory Utility Labs
The umbrella imprint the work is published under. Not a product — the name on the publications. Its charter is still being written; the work is not.
Read about the Labs →Ektar
The operating methodology behind everything here: every consequential claim carries a status — verified, inferred, guessed, unknown.
Read about Ektar →Web Design
Portfolio work including K&A Doors. Publication pending the client's written consent; the page exists as a draft only.
Flagships
Published and documented. Nothing here claims users, adoption, or results the repositories do not show.
Interlock — v0.1.0, MIT, CI green
Zero-dependency Python safety substrate for AI agents. Expiring authority leases, a global e-stop, a fault bus, provenance-stamped claims. The rare agent-safety project that ships a release, a changelog, and green CI on day one.
View on GitHub →AEGIS — v1.2.0, MIT
A local-first agent operating system for AI coding work. Receipt-backed token ledger, honest-yield rule (no savings claim without an admitted proof pair), frozen /v1 API.
View on GitHub →Reclamation Evidence Ledger — MIT
Independent satellite watchdog for Alberta's orphan wells. Free Sentinel-2 imagery, automated change detection over a 27-site pilot, 21,892-site registry, provenance-stamped evidence packets.
View on GitHub →Star Lab — MIT, CI green
A free local agent research lab: 14 offline-first modules for experiments, model evaluation, and mission control. Commissioned and directed; built with Grok as a fair head-to-head stage for coding agents.
View on GitHub →Sovereign Contracts — MIT
Local-first, auditable smart-contract pipeline — Hardhat 2 + Python, gated compile/test/coverage/deploy, 100% coverage floor on production vaults. No cloud accounts required.
View on GitHub →Hermes — Apache-2.0
Fail-closed macOS execution layer for agent-driven work. Allowlisted capabilities, exact target resolution, independent validation, redacted append-only receipts. Architecture and threat model published; implementation in progress.
View on GitHub →Why this portfolio stands out
- Ships releases, not just READMEs. Interlock v0.1.0 and AEGIS v1.2.0 are tagged and released — most comparable solo projects never cut a release.
- Honest by construction. Claim tiers on every project; the honest-yield rule refuses to fake efficiency numbers. In a field full of inflated benchmarks, restraint is a competitive edge.
- Local-first as doctrine. No cloud lock-in, no API-key sprawl — everything runs on the machine, ledgered.
- Safety as substrate, not patch. Interlock bakes leases, e-stop, and fault buses into the agent loop — the way Boston Dynamics bakes them into robots.
What is not claimed: users, adoption, stars, or benchmark wins. All six projects are new (published August–September 2026). New and undiscovered is fine — the work is the argument.
Contact
The fastest way to reach me is email: marcus@marcusrichards.dev.