I build and secure agentic AI systems for the enterprise, and I sit on both sides of that conversation: the architecture that has to work, and the business case that has to hold. Those two jobs usually belong to different people, which is most of why AI programs stall.
Most organizations haven’t scaled AI past a pilot, and the reasons are predictable. I work embedded with client teams, holding both the technical design and the business argument: the pairing that keeps a program alive past its first review.
ROI and value models that hold up under a CFO’s scrutiny, defended in the room where most value cases come apart.
Agents built hands-on, scoped to the user’s permissions, fenced against injected instructions, and gated by human approval.
Large security programs for enterprise clients, with security as a design input instead of a retrofit two weeks before launch.
One working deployment beats a three-year roadmap. I ship the workflow that pays for itself fastest, against the systems you already run.
A demo agent is a model and a few tools. An agent allowed near production systems is a distributed system with an untrusted natural-language interface. This is the reference architecture I build to, with each boundary mapped to the threat it exists to stop.
The boundaries do the work. Every natural-language input arrives untrusted, including retrieved documents and messages from other agents. The agent carries the calling user’s permissions and no more. High-impact actions stop for a person. And everything is logged well enough to reconstruct afterward what the agent did and why it decided to.
Mixture-of-experts models route each token to a handful of specialists, invisibly. Latency and failure patterns leak that routing anyway. Probe long enough and you can map which expert handles what, then target the weakest safety behavior. Reconnaissance rather than exploit, which is why it falls off threat models.
Provisioning, rotation, and deprovisioning all assume a human on the other end. Give a thousand agents scoped credentials and the directory quietly becomes the largest unmanaged identity surface in the company. Each holds real authority, and nobody has built a process for offboarding one.
Most evaluation catches single-turn failures. A corrupted memory or retrieved document persists instead, shaping decisions days later with no obvious link back to the injection that caused them. That distance between where harm surfaces and where it entered is the whole problem.
An agent’s capability surface is now effectively a package manifest. Every connected tool is code you did not write, executing with your users’ authority against your real systems, and the ecosystem ships far faster than any review process has managed to keep up with.
Innovative safety system for autonomous vehicles designed to enhance occupant protection through advanced robotic mechanisms and intelligent response systems.
View Patent →Novel methodology for generating and simulating complex traffic scenarios to improve autonomous vehicle testing and validation processes.
View Patent →Bio-inspired approaches to helmet design combining finite element modeling and experimental testing. Prototype keeps acceleration peaks under ECE22.05 Head-Injury-Criterion limits.
Conference → Publication →Lebanon's 2019 currency crisis, one of the world's worst financial collapses since the 19th century, threatened the country's entire economic and social fabric. Filmed during my time at Michigan Ross, I explore the twin deficit phenomenon at both the fiscal and balance-of-trade levels.
Watch on YouTube →Continuum damage coupled plasticity model and validation for magnesium alloy AZ31B with time-dependent and anisotropic behavior.
Read Publication →Finite element and experimental evaluation framework for safer helmet design with improved energy absorption and impact mitigation.
Read Publication →
Writing the code stopped being the bottleneck. Knowing what to ask for, and recognizing when what comes back is confidently wrong, is the job now.
I’m a Managing Consultant at IBM working on enterprise AI and cybersecurity. I run large programs
for major clients, and I build the agent systems myself. Those two things almost never live in the same
person, and the gap between them is where most AI investment quietly dies: architecture nobody can
justify, or business cases nobody can build.
The agentic shift changed what technical work rewards. For a decade the prized engineer was the
meticulous one, careful with every line. That skill is now table stakes. The work today is decomposition,
delegation, and verification: framing a problem precisely enough that a system can act on it, engineering
the context it needs to act well, and holding a high enough bar to catch output that is fluent and wrong.
Those are management skills. The people adapting fastest are the ones who already knew how to set a
brief, delegate it, and inspect what came back.
I’ve spent years translating between the engineers who build systems and the executives who fund
them, which turns out to be the same shape as directing an agent: set the goal, bound the authority,
judge the result against what the business actually needed. I’ve gone deepest in automotive,
manufacturing, and energy & utilities, and the agentic work keeps pulling me wider, into e-commerce
and wherever else an enterprise is putting agents near real systems. The problems repeat; the industry
changes the constraints. I hold an MBA from Michigan Ross in strategy and finance,
IBM’s Advanced credential in Generative & Agentic AI, and Gold-level industry certification
in Automotive.
I work both ends of the client relationship: winning the work, then delivering it. Consultants usually specialize in one or the other.
Led an 8+ person team to launch Ford’s first ADAS cloud analytics platform and automated scenario extraction, scaling L2+ simulation throughput on a Kubernetes-based framework. Earlier, pioneered a scenario-based MBSE method for SOTIF verification, advanced data governance for ML assets, and contributed to SAE J3187 standardization.
Whether you’re getting a first deployment funded or securing agents that already touch production, the useful next step is the same: a short conversation about where it stands and where it’s stuck.
Let’s Talk