I don't demo agents. I build agentic systems for enterprise.
Data is the through-line of my work. I design and run coordinator-and-specialist agent systems with guardrails built in, and operate them against real stakes: a car-rental business whose back office runs on a team of agents I built, live markets through read-only samplers, and production data pipelines. The interesting part is not what an agent can do, it is what it is allowed to do.
Claude Certified Architect(Foundations) · AnthropicOpen to: AI and agent engineering, system design, technical program management, and data-collection programs for autonomous vehicles and robotics.
The safety floor
Anyone can make an agent act. The engineering is making it stop. Every system I build runs under a permission model: routine work is allowed, consequential work waits for a human, and a few things can never happen at all.
Permission modelIllustrative
Sample actions an agent might attempt, and the verdict each one gets. Allowed runs on its own, Needs approval pauses for a human, Blocked is a floor the system cannot cross. Click any row to see why.
An illustration of the guardrail model my systems run under, not a live or connected system. No real action is taken here.
Draft a weekly report Allowed
Why it's allowedThis only reads and summarizes information the agent already has. Nothing is sent, changed, or committed, so a mistake costs a re-run, not a real consequence.
Read fleet telemetry Allowed
Why it's allowedReading data leaves nothing changed and nothing to undo. There's no action here for a human to approve, since the state of the world doesn't move.
Send a customer email Needs approval
Why it waitsThis leaves the system in the business's name and lands with someone the business doesn't control. Tone, accuracy, and timing are judgment calls, so a person reads it first.
Push to production Needs approval
Why it waitsThis changes what real users see next. It's usually reversible with a rollback, but the blast radius is big enough that a second set of eyes checks it before it ships.
Move funds or place a trade Blocked: safety floor
Why it's blockedMoney that moves doesn't un-move the way a bad draft or a bad deploy does. No approval step makes that safe enough, so there's no path for an agent to even request it.
Delete records or change access Blocked: safety floor
Why it's blockedDeleting data or changing who can see what can destroy information or open a door that can't be quietly closed again. That sits below the approval line, not behind it.
Agent-driven delivery
The most honest proof of the approach is the site you are reading. It was designed, built, rebuilt, tested, and deployed by orchestrating my own agents, with a human keeping the decisions and the agents on the legwork.
An ideation think-tank proposes directions in parallel, specialist build agents implement, an auditor gate reviews every wave and returns pass, warn, or block, and persona-based testers stress the result. The loop runs to a written definition of done. One operator covers ground that traditionally spans several roles: design, front-end build, QA, accessibility review, and copy.
- Parallel specialistsIndependent work runs at the same time instead of one model doing everything in sequence.
- Auditor-gated wavesEvery wave is reviewed and returns pass, warn, or block before the next one starts.
- Persona testing built inPersona-based testers click through the result and loop findings back until it meets the bar.
- One-operator leverageA single person orchestrates the work that usually needs a small multi-role team.
EstimateWhat this approach saves, assumptions shown
A hand-built equivalent at this level of polish, interactivity, accessibility, and QA is realistically one to two weeks of focused solo work, or a small multi-person sprint. Orchestrating agents compressed that to about two days of part-time work, with the build-and-QA loop running in hours rather than weeks.
Assumptions: a freelance web rate of roughly $75 to $150 per hour, and a one-to-two-week hand-built baseline for the same scope. On those assumptions the saved effort lands on the order of several thousand dollars, delivered by one operator. This is an estimate with its assumptions shown, not a measured invoice or a stopwatch result. Ranges are deliberate.
Selected work
The same through-line runs through every system: state the capability, then state the leash. The production work runs where mistakes carry real costs; the open-source systems are public and inspectable.
Production agent systems in a high-stakes environment
Production agentic systems with guardrails: multi-agent work where capability is bounded by hard safety limits, with a human on every consequential decision.
Real stakes, bounded by designVantage OS
A Claude Code plugin where a coordinator routes every request to the right skill or sub-agent, then gates real-world actions through a QA filter and a permissions-tier system before they reach the human.
10 skills, 8 agents, publicLocal-first practice assistant
A Windows desktop app for a solo therapist, built on the Agent SDK, that handles the back office while the sensitive data stays encrypted on the device. Open-source and anonymized.
Privacy is the architectureBuilt in production
I design and run production agentic systems in a high-stakes environment: multi-agent work where capability is bounded by hard safety limits, with a human on every consequential decision and the agents on the legwork.
The open-source systems here show the same rigor in code you can inspect. More on the production work →
Work with me
If you are hiring for agent engineering, system design, or technical program management, I am glad to talk. The fastest path is email, and the code is on GitHub.
Open to: AI and agent engineering, system design, technical program management, and data-collection programs for autonomous vehicles and robotics.