An agentic AI engineer in West Palm Beach who runs agents in production, not demos
Agentic AI is easy to demo and hard to operate. I design multi-agent systems that do real work — research, writing, outreach, monitoring, publishing — inside guardrails: typed tools, approval boundaries, ledgers, and checks that prove an outcome happened rather than that a job ran.
What agentic engineering means in practice
An agent is a model with tools and a loop. Making one useful means deciding what it may touch, what it must ask before doing, how its actions are recorded, and how you find out when it is wrong. Most of the engineering is in those boundaries, not in the prompt.
The systems I run use typed, permission-scoped tools instead of open shells, keep an action ledger for every outbound step, gate anything customer-facing behind deterministic quality checks, and maintain a registry of assertions — one per past failure — that runs every day so nothing regresses silently.
What I engineer
Multi-agent orchestration
Operators with scoped tools, schedules, caps and receipts — coordinated, not chaotic.
Tool and permission design
Typed tools, allow-lists and approval boundaries instead of an open terminal.
Verification layers
Deterministic gates, live checks and a never-twice assertion registry.
LLM integration
Retrieval, structured outputs, provider fallbacks and cost control across models.
Observability
Ledgers, health checks and plain-English alerts so humans stay in the loop without being in the way.
Reliability engineering
Timeouts, idempotency, dedupe and rollback — the boring parts that keep agents trustworthy.
Production examples
A fleet of operator agents that manufactures demand for several businesses: research, outreach drafting, call handling, follow-up and reporting, each within caps and a ledger. An autonomous content system that researches, writes, gates, publishes and verifies articles across fifteen websites and audits its own output against a written standard daily. A consolidated owner inbox built so no inbound is missed.
How I work
- Outcomes over activity: every automated step must produce a receipt that can be checked.
- Fail closed. When a check cannot run, the system stops, it does not guess.
- Small typed changes, tested, reviewed adversarially before they ship.
- Every incident becomes an assertion that runs daily — the same mistake is not made twice.
- Based in West Palm Beach, working with companies across Florida and nationwide.
Frequently asked questions
What is an agentic AI system?
Software where a language model plans and executes multi-step work using tools — reading data, calling APIs, writing, sending — inside boundaries an engineer defines. The engineering is mostly in those boundaries.
Which models and frameworks do you use?
Whatever fits the job: frontier models from several providers with fallbacks, typed tool interfaces, and orchestration I build rather than a single vendor's framework. The design is provider-agnostic.
How do you keep agents from doing something harmful?
Scoped permissions, approval gates for anything external or irreversible, action ledgers, caps, and daily assertions that catch regressions. Customer-facing outputs pass deterministic quality gates before they ship.
Do you work with companies outside Palm Beach?
Yes. I'm based in West Palm Beach and work remotely with teams across Florida and the United States.
Can you take over an existing AI project?
Often. The first step is an audit of what the current system actually does versus what its reports claim — that gap is usually where the risk is.
Bring me the agent that almost works
If you have an AI workflow that demos well and fails in production, that is the exact problem I solve.
Start a project