A Memory Utility Publication · Portfolio Vol. I
AI Systems Operator & Automation Architect — Calgary, Alberta
I build AI systems that are safe, economical, and auditable on ordinary hardware. Every project below ships with its evidence attached — what is proven, what is in progress, and what is not claimed.
The portfolio has a spine now. Interlock is the mechanism — authority that expires. Exhibit is the intelligence — every claim presented as evidence. Reclamation is the proof applied — accountability measured from orbit. Each dossier below is a full reading: the problem, the mechanism, the record, and what is not claimed.
Agent safety substrate
Authority is a loan, not a possession: leases that expire, a global e-stop, a fault bus, and claims that carry their provenance. Zero-dependency Python, v0.1.0, MIT, CI green — the tests ship with the release.
What it proves: safety can be a runtime property, not a policy document.
Read the dossier →
Open-source intelligence, presented as evidence
Defensive OSINT for infrastructure owners: 11 public sources fused, a certificate-transparency radar, and a temporal sentinel that remembers what it couldn't see. Born from a postmortem — the pipeline once believed a re-served scan at 90% confidence. Now the collector never verifies; a human disposes.
What it proves: intelligence earns trust the way testimony does — provenance, or it didn't happen.
Read the dossier →
Environmental accountability, from orbit
Sentinel-2 change detection over Alberta's oil and gas footprint — 99 assessed sites across 13 licensees, each with a machine-readable evidence packet on a live public dashboard. Its best feature is a mistake it caught in itself: the first method measured the calendar, not the land, and the audit rebuilt it.
What it proves: public satellite data can hold remediation to account — and honest methods survive their own audits.
Read the dossier →Autonomous systems fail in three predictable ways: they exceed their authority, they burn budget nobody measured, and nobody can reconstruct what they did. My work attacks all three — with runtime controls, token governance, and ledgers that stamp every claim with its provenance.
Bounded authority at runtime: leases that expire, an e-stop that actually stops, a fault bus that records. Interlock and Hermes treat an agent's power as a loan, not a possession.
Tokens are finite inventory. AEGIS packs context once, reuses it while the bytes are unchanged, and holds a protective reserve floor — with an honest-yield rule that keeps savings claims null until an admitted pair of runs proves them.
Every consequential claim carries its status — verified, inferred, guessed, unknown. Ledgers, receipts, and evidence packets make the work reconstructible by someone who was never in the room.
The dossiers above are the argument; these are the instruments behind it — the governance substrate under trial, the lab where doctrines get tested, and architectures still on the drawing board. Each labeled exactly for what it is.
Governance substrate, under trial
Token governance for long coding sessions: pack once, reuse while the bytes are unchanged, land, ledger. A protective reserve floor of ≥80% under a weekly budget keeps sessions honest, and the honest-yield rule refuses to print a savings percentage until an admitted pair of runs earns it. A supporting thesis to the flagships — the trial is the point.
What it proves: economy can be engineered — and measured without lying.
Read the volume →
A free local agent research lab
Commissioned and directed by me, built with Grok: a local-first laboratory for running agent experiments on ordinary hardware — no cloud bill, no data leaving the bench. The workshop where doctrines get tested before they get published.
What it proves: serious agent research doesn't require serious infrastructure.
Inspect the repository →
Smart contracts, auditable by design
A local-first pipeline for sovereign contract work — every stage inspectable, every artifact reproducible, nothing hidden behind a hosted black box. Built for contexts where the contract is the relationship and the audit trail is the product.
What it proves: on-chain work can carry its own evidence.
Inspect the repository →
Controlled execution for macOS agents
An architecture for bounded agent execution on the Mac: allowlists decide what may run, gates decide when, receipts record what did. Currently at the architecture and design stage — the drawings are honest about being drawings.
What it will prove: the desktop agent can be governed like infrastructure.
Inspect the repository →“A savings claim stays null until an admitted pair of runs proves it. No admitted pair, no percentage. The rule is the feature.”
Pack once, reuse while the bytes are unchanged. Context is inventory; inventory gets counted.
Protective reserve floor of ≥80% under a weekly token budget — austere by design, aggressive on drift.
Land every change to the ledger. The 30-day existence proof for the first honest-yield pair is in progress; until it lands, there is no number here.
The Ektar standard: the finest grain, no fog. Nothing on this site is louder than its evidence. Consequential claims are tiered at the point of writing — and the tiers are part of the design, not the fine print.
Safe figures only — each one traceable to a repository, a release, or a ledger. The boundary below is load-bearing: it is what makes the numbers above believable.
The publishing imprint behind the work — Calgary, Alberta. Technical monographs, working systems, and the ledgers that keep them honest. Introducing AEGIS, Technical Specifications Vol. I, is its first volume.
The working standard: provenance, uncertainty, time, decisions, and change — preserved as carefully as the work itself. The voice of these pages.
For roles, collaborations, and serious inquiries about the work. Every message gets an answer; every claim in the reply carries its status.