$4M portfolio
Owned a 30-application estate and supported the retirement of 80+ servers, virtual machines, and databases.
About
AI operations that have to work in the real world.
Twenty years of turning ambiguous problems into measurable outcomes — enterprise product and operations leadership on one side, hands-on research and building on the other. The two are not separate careers. Each one keeps the other honest.
$30M in programs
Technology pillar lead across a manufacturing consolidation and an international agricultural build.
Three continents
Led distributed technology analysts across the United States, India, and the Netherlands.
Enterprise
Operating inside constraints, not around them.
My enterprise work has been in manufacturing, agriculture, and records management — environments where systems are old, dependencies are tangled, regulators are real, and replacing the platform is not on the table. I led the technology pillar on a $10M manufacturing consolidation moving 1,500 finished goods across six production lines, and on a $20M international agricultural build that required restricted-data handling and synchronization across multiple ERP systems.
That work is where I learned the discipline the rest of this site is built on: a change is not finished when it ships, it is finished when someone can audit what happened and reverse it if they were wrong.
Independent
Building the thing, not just specifying it.
Through my own research studio I have delivered roughly ten engagements across blockchain, fabrication, and security testing, and I build continuously: purpose-built MCP servers for agent tooling, a local-first blockchain forensics workstation, an evaluation platform for private security models running on a single consumer GPU, and applied research into low-bit retrieval for constrained infrastructure.
I keep this hands-on for a specific reason. It is difficult to govern an AI system, scope a security review, or judge an engineering estimate honestly if you have not recently been the person who had to make it actually run.
How I work
The hard middle.
My focus is the gap between a promising technical capability and an operating model that has owners, evidence, controls, measurable outcomes, and a path to scale. In practice that means being specific about maturity — production, pilot, prototype, or research — and stating limits in the product rather than in a disclaimer nobody reads.
It also means being willing to lose the good number. The research case study on this site exists because a fair comparison erased our own headline result, and publishing the smaller, true version was the right call.
Education
Background.
M.S. in Organizational Leadership, magna cum laude. B.S. in Criminal Justice Studies. Currently pursuing an M.S. focused on cybercrime — which is where the investigative and forensics work on this site comes from, and where it is heading next.
Portfolio
Projects, documented as I go.
Each case study describes decisions, methods, and verified outcomes. Client work appears only as an anonymized write-up, reviewed before it goes live. Names, figures, and screenshots from client systems are never shown.
If you would rather start from how I work than from what I built, the ten rules these projects run on are written up separately — each one linked to the case study where it was paid for.
Governed automationAgents / Teams15 scoped agents, 6 teams that call them.
GovernanceReviewSecurityDiscovery
AgentsEach one narrow, report-only unless stated, and cheap to run.
signal-chief
Reads current state and returns one ordered list of what matters next.
verification-gate
Runs a repository's declared checks and returns an evidence packet before anything ships.
whitehat
Report-only security review of owned assets. Findings first, severity-ranked.
citation-audit
Re-derives every citation against its actual source rather than reading it back — including the self-describing numbers on a live public page.
eval-referee
Accepts that a measured number is real and asks what it is allowed to mean: control, held-out set, fair comparison, sample size, harness.
approval-triage
Independent second read on a single risky pending action, before it reaches a human for approval.
registry-sync
Diffs declared configuration against what is actually on disk, and flags the gap.
staleness-sentinel
Drift check across agents, documentation, backlog, and operating rules.
governance-sentinel
Tracks backlog aging and a cost-versus-output read over time.
incident-rca
Builds a redacted evidence packet, timeline, and classification after an incident.
incident-rca-reviewer
Independent challenge pass on an incident draft before it is accepted.
claim-debate-reviewer
Four-persona adversarial debate to pressure-test one contested claim.
vault-scout
Content discovery sweep for duplicate work and reusable material.
night-watch
Quiet, low-noise health pass over workspace configuration and sync state.
overwatch-ux-reviewer
Reviews interface changes for flow, small-screen layout, and empty or error states.
Agentic teamsComposed calls. The shape is the point, not the headcount.
Drift sweepin parallel
staleness-sentinelregistry-syncgovernance-sentinel
Three read-only passes run in parallel and land as one answer to “has anything quietly moved since we last looked?”
Incident reviewin order
incident-rcaincident-rca-reviewer
Evidence first, then an independent challenge. The agent that gathers the evidence is never the one that clears it.
Ship gatein order
verification-gatewhitehat
Declared checks, then a security pass. Anything public-facing clears both or it does not ship.
Publish gatein order
citation-auditeval-refereewhitehatsensitivity sweep
A page can be secure, tested, and still be wrong. True first, then entitled to its framing, then safe, then free of internal context — and it runs before the publish, because a gate after the deploy is a post-mortem.
Repair teamin order
parallel auditssingle writervalidation
Audits fan out read-only; exactly one stage is permitted to write, and its work is validated before it counts.
Weekly sweepin parallel
driftregistrycontentgovernance
Runs unattended on a schedule and stays silent unless something actually needs attention.
AI enablement · agent governance
Running agent teams that cannot authorize themselves
Fifteen role-scoped agents in daily operation, none of which can write a single file. The capability lives in what they are allowed to call; the safety lives in what the call surface does not expose.
Read case study →Agents · on-chain licensing
Deploying agents for blockchain autonomy
Agents run an entire software licensing pipeline on a live blockchain. The token is not a receipt — it is the key to an MCP app handoff and every update that follows, for as long as the holder still owns it.
Read case study →Applied AI research · private inferenceHead-to-head complete · held-out validated
Building the measurement that argued against the modification I wanted
A security-model evaluation platform on one 12 GiB GPU, built to decide whether to strip a model's safety refusals. Three models were measured across four capability axes, the headline findings were re-measured on held-out items, and the evidence argued against the modification three independent ways. The measurement was commissioned to justify a decision and it declined to.
Read case study →Applied AI research · evaluation methodPromoted, then un-promoted
A runtime configuration I promoted, and the held-out run that took it back
Category-routed reasoning beat always-on reasoning on both accuracy and latency — 16 of 16 at 30 seconds against 15 of 16 at nearly two minutes. It was promoted as the champion runtime. A held-out suite run the same day reversed the result.
Read case study →Applied AI researchPaper
Compact retrieval as infrastructure repurposing — and the headline I had to retract
A deployment-fit study of low-bit retrieval indexes, in which the fair comparison erased our own best number and produced a narrower, more defensible finding.
Read case study →Investigative toolingv0.4.0 · Gate 1 signed
Building a forensics tool that refuses to guess
A local-first blockchain investigation workstation designed so that its evidence log can survive a challenge — and so that the things it cannot prove are stated in the product, not the disclaimer.
Read case study →Smart contracts · irreversible systems
Shipping smart contracts to a public network, where you cannot patch a mistake
Designing, reviewing, and deploying randomness-dependent contracts to a live blockchain — and treating an oracle that never answers as a design requirement rather than an edge case.
Read case study →Agent tooling · secure delivery
Purpose-built agent tooling, and getting software into someone else's hands safely
Three MCP servers that give AI agents real capability inside hard authority limits, and a delivery flow that assumes the recipient should inspect before they run.
Read case study →Enterprise operations · portfolio modernization
Retiring eighty-plus systems from an estate nobody could afford to stop using
A 30-application portfolio carrying more than $4M in annual operating expense, and eighty-plus servers, virtual machines and databases retired from under it — sequenced so that every step stayed auditable and reversible until the last one.
Read case study →