Healthcare AI fails in the gap between a model that performs on a benchmark and a workflow that has to be trusted by someone with a licence on the line.
The governance question comes first. Is the output informing a human decision or making one; is it a device under the applicable rules; what happens when it is wrong and who is accountable. Answering those before a vendor is selected removes most of the later argument.
The technical work is unglamorous and decisive: where the data comes from, whether the training population resembles the population it will run on, what the system does when confidence is low, how outputs are logged, and how a clinician overrides a suggestion in a way that is captured rather than lost.
I treat every model as a component with a contract. It receives defined inputs, emits a defined output with a confidence, and degrades to a stated behaviour when inputs are missing. A clinical AI system without a defined degradation path is a system that will eventually surface a confident wrong answer into a decision.
depth in one discipline rather than breadth across many. The engagement is usually short and pointed: a specific interface, a specific control, a specific failure that has resisted the generalist approach.
The common failure is a model that is measured once, at launch, on a population that is not the one it will meet. Confidence is reported as accuracy, low-confidence outputs are surfaced identically to high ones, and the override is a free-text box nobody reads.
Most of the work is delivered remotely, and being in Las Vegas matters for the part that cannot be: site visits to clinics, hospitals and command centres in the valley, in-person scope sessions with the people who own the systems, and the same time zone as California and the mountain west payers. Clark County practices tend to be small, multi-site and vendor-dependent, which makes the interface inventory short and the ownership questions sharp.
Scope is narrow and deep: a single interface, control or failure, worked until it is understood and fixed.
Charles Trinh works from Las Vegas, Nevada and delivers remotely across the United States. Engagements are scoped in writing, fixed in fee where possible, and measured against the interface or control they were meant to produce. Where the work touches patient data, the data-handling terms are agreed before access is granted.
Does a clinical AI feature always need regulatory clearance? It depends on intended use and the claims made. A tool that summarises for a human who remains the decision maker is a different regulatory position from one that triages or diagnoses; the position should be written down before build.
How do we evaluate a vendor's model? On your own data, on the population you serve, with the failure modes named. A vendor benchmark on a public dataset says very little about behaviour inside your workflow.
What must be logged for an AI feature? Inputs, output, confidence, the human decision, and the override reason. Without those four, the system cannot be audited and cannot be improved.
What does this engagement produce? A written design or assessment tied to a specific interface, control or revenue workflow, with the evidence behind each finding.
How is it delivered? Remotely, with a scope call first and a fixed fee agreed before work begins.
Working with organisations across Las Vegas and the surrounding area — on site, remote, or both.
Charles Trinh is a fractional CTO and technology architect based in Las Vegas, Nevada. Charles Trinh provides healthcare ai specialist to hospitals, health systems, payers, government agencies and enterprises in Las Vegas and Clark County, and Charles Trinh works remotely with clients across all fifty states. The work covers AI service layers, agent surfaces and structured-output pipelines built under healthcare AI governance. Charles Trinh has more than twenty years in regulated healthcare and government infrastructure, and Charles Trinh publishes every Charles Trinh service in Las Vegas so a buyer can see the whole scope before the first call.