AI

Charles Trinh — AI Model Governance in Healthcare | Las Vegas, NV · Las Vegas, Nevada · Henderson · Clark County, Nevada · North Las Vegas, Nevada

A model inventory table with owner, version, review date and approval state
Charles Trinh — AI Model Governance in Healthcare

Governance is the set of controls that makes an AI deployment answerable. Without it, every question about a model output starts from a blank page. I build the inventory, the approval trail and the change control that end that.

Write the inventory before the policy

Most governance programmes start with a policy document. I start with an inventory of every model already in use, including the spreadsheet scoring tool that a department has quietly relied on for a decade. A policy governs nothing that is not listed, and an organisation that cannot list its models cannot answer a regulator's first question. The inventory is the foundation, and building it frequently surfaces deployments nobody knew existed, from a vendor module embedded in billing software to a local script a researcher runs on an unpaid colleague's laptop.

The inventory entry records the model identity, the owner by name, the purpose, the data it reads, the decisions it influences, the current version and the last review date. Purpose is the field that determines everything downstream, because the same model with a different purpose is a different governance problem. A model that flags a patient as high risk for a social-work conversation is governed differently from the same model deciding whether a claim is paid out.

The inventory must be machine-readable, or it collapses. I build it as a table in the organisation's database, with a schema and an import path, so that adding a model is a routine change request and reviewing all models due is a query. A spreadsheet inventory becomes a stale document within a quarter. That is the standard failure and it is avoidable from the first day.

The approval trail and intended use

An approval trail is a sequence of named decisions with evidence. Each model entry has a state that moves through draft, evaluated, approved for a specific use and withdrawn. A transition records who made it, what evidence was cited and the date. A model whose state cannot be traced from draft to approval is not approved in any meaningful sense, even if a committee once discussed it and someone has a note somewhere.

The intended use statement is the pivot. It names the decision the model influences, the population, the setting, the human role, and what the model is explicitly not for. It is what distinguishes a regulated medical device from an internal analytical tool, and I write it before building the evaluation, because the evaluation's pass criteria come from it. The statement should read like an instruction manual section, not a marketing description, and it should be short enough that the committee can actually discuss it.

The same database that holds the inventory and the approval trail holds the evaluation references. A model record links to its evaluation report, its stratified results, its open issues and the monitoring dashboard. That makes the governance review a review of data rather than a review of slides, and it makes the next quarterly review fast because everything is already assembled.

Change control for model versions

A hosted model is not a fixed asset. Vendors ship version changes silently, retrained models on schedules their customers do not fully control, and occasionally change a routing policy or a safety filter that alters outputs without any new version number. Governance that assumes the model does not change is governance of yesterday's system. I build change control around the fact that the system changes continuously and the approval must therefore be a standing gate rather than a one-time event.

Every model call is tagged with its version, hyperparameters and any inference parameters that matter. When a model version changes, the change is recorded automatically and a comparison run fires against the frozen reference corpus. That run determines whether the new version needs re-evaluation or just re-recording. Re-evaluation is not the default for every minor update, but every update is recorded, and the decision to skip re-evaluation is itself a recorded decision with a named owner.

The gate has to be in the pipeline, not in a document. If the conformance run that blocks a code deployment does not block a model version change, the model version change is not really governed. I prefer building model swaps so that they flow through the same release gate as code, with a comparison report attached. That is the arrangement that keeps governance honest after the first compliance review is over.

The risk register and stratified evidence

A healthcare model risk register is not a list of technical risks; it ranks ways model output could harm a patient, a group or the organisation by likelihood and impact, as a clinical risk register would. Discrimination in a subgroup, silent degradation after a population shift, over-reliance by staff and misrouting of a case needing review are the entries I see most, and each needs a named control in the register.

The register connects to the evaluation. A discriminated-population risk needs stratified results, which are a fixture on the evaluation report, not a special request. A degradation risk needs a drift monitor and a re-evaluation trigger. An over-reliance risk needs interface design and training described in the model record. The register is the control book; the model records are the working records that cite it. When the two disagree, somebody has updated one and not the other, and the update procedure usually needs fixing.

Stratification deserves its own measurement. For each model I list the populations where the intended use applies and I report results separately for each. Where a population is too small to evaluate, I say so and I say what the risk register does about it. A model that is safe on average but untested in a group that receives it is a model that has governed technical performance and not clinical risk, which is the risk that matters most.

Monitoring, drift and a scheduled re-evaluation

Monitoring for a healthcare model is a set of measurements tied to the intended use: input distribution, output distribution, downstream action distribution and, where a clinician records it, agreement between the model output and the clinician's decision. The control chart that alarms is set by the governance group from the evaluation data, not left at a software default. This is the design in observability-synthetic-monitoring applied to a model rather than a service, and it uses the same alerting discipline.

Drift shows up in input distributions before it shows up in outcomes, so weekly distribution checks catch a shift early. But drift alone is not a decision. The governance record links each drift signal to an owner and a decision — retrain, recalibrate, withdraw, or continue — and the decision is recorded with the data. What I want to avoid is a dashboard that shows a shift nobody acts on for months, and a model that is quietly worse because of it.

A scheduled re-evaluation closes the loop. I set a cadence — annually for stable models and quarterly where the population mix changes fast — and the inventory itself triggers the review by surfacing all models due in the next ninety days. Re-evaluation uses the same harness, produces a new report that supersedes the old one, and records the reason the result is different. The history is then visible to anyone who asks how the model's performance changed.

Evidence and retention for the external audit

Governance is judged by what an external reviewer can retrieve in a reasonable time. The organisation I work with should be able to answer, within an hour and without a developer, every question a regulator, accreditation surveyor or internal audit might reasonably ask about a given model. The inventory, the approval trail, the evaluation report, the change log and the monitoring history should be one query away and readable without engineering help.

Retention follows the use. Where model output entered a clinical decision, the record lives as long as the chart, and where output entered a billing or administrative record, it lives as long as that record's retention rule. I avoid indefinite retention of raw inputs and treat them as phantoms; the analysis artifacts and scores are kept, the raw input is not, unless the use requires it. That is a privacy decision as much as a governance one and it is made with the privacy officer.

The deliverable is the inventory schema, the approval-trail tooling, the change-control gate and the re-evaluation schedule, with a quarterly evidence pack that closes each review. My background includes building and operating backend data systems in HIPAA-adjacent contexts; I do not hold a medical degree and I will not sign off on clinical judgement. The clinical sign-off on intended use and stratification lives with your clinical leadership, and that is where it belongs.

What an engagement looks like

The first month is inventory building. That is unglamorous work and it often reveals that the organisation has more models than it thought and fewer owners than models. I build the schema, migrate the existing spreadsheet where one exists and interview the department heads who know what is really running. The output is the inventory, populated, and a shortlist of models that need an approval trail retroactively completed.

The next phase is the approval trail and the change-control gate, phase-built around the models that matter for regulatory or clinical risk. That is the work where the governance becomes real. Once the gate is in place and wired into the pipeline for one or two models, extending it to the rest is a matter of registration rather than new engineering, so the cost per model declines sharply as the programme grows.

None of this amounts to an accreditation, and I do not represent that it does. It is a set of processes and a set of records that an auditor can inspect without engineering help. Where third-party attestation of governance is required, that is work for a separate party and I say so before the engagement begins. What I build is the working record the attestation will be built on.

Proof

  • ONC (g)(10) self-test PASS, 317 of 317 subtests, April 2026 — public suite URL
  • 67 public hostnames, 51 serving, one AWS account
  • 104 clearinghouse enrolments across 53 jurisdictions
  • USCDI v7 proposal — Patient Digital Identity Credential — listed at isp.healthit.gov
  • SAM.gov UEI MVKQW9DKKGB8 · CAGE 20KM5 · NPI 1194645564
  • AWS Business Associate Addendum active since 18 August 2026

Where this work is delivered

Engagements run from Las Vegas and reach across the valley: clients in Las Vegas, NV; Las Vegas, Nevada; Henderson; Clark County, Nevada; and North Las Vegas, Nevada — plus remote delivery nationwide.

How an engagement runs

1. Trace one real case

Follow a single patient or claim end to end with a timestamp at every hop, before any document is written.

2. Written findings

A findings list naming the defect, the system, the owner and the date.

3. Monthly review

Review the list monthly and check whether the dates held.

Frequently asked

What is this service?

This page describes charles trinh — ai model governance in healthcare — the scope, the failure modes it prevents and how the work is measured. It sits in the AI and Automation category at trinh.io.

Where is the work delivered?

Engagements run from Las Vegas, Nevada across the valley — Las Vegas, Henderson, North Las Vegas, Spring Valley, Paradise, Winchester, Sunrise Manor, Enterprise, Summerlin, Clark County — and remotely anywhere in the United States. On-site work in the Las Vegas valley, remote delivery elsewhere.

What does it cost?

Standard rate $195 per hour, committed $165 per hour at twenty hours or more per week for three months or more; $1,600 standard day, $1,350 committed day. Scoped in writing before work starts, fixed fee where the deliverable is defined.

Who is Charles Trinh?

A technology architect in Las Vegas, Nevada, working in healthcare identity, consent and interoperability. Not a physician, not a medical practice, and not affiliated with any other Charles Trinh.

How does an engagement start?

With one trace of a real case and a written findings list naming the defect, the system, the owner and the date. The first call is twenty minutes and free: 702-559-3489.

Engagement

Scoped in writing before work starts; fixed fee where the deliverable is defined. First call is 20 minutes and free. 702-559-3489 · charles@trinh.io

Talk to Charles

Record

Profile

Healthcare technology architect working where federal regulation becomes production infrastructure. Built and operates a FHIR R4 clinical platform — SMART App Launch authorization server, patient and provider portals, SQLite — on a hardened single-host AWS estate of 67 public hostnames. The platform returned a self-test PASS on the ONC (g)(10) Standardized API test suite in April 2026, 317 of 317 subtests, publicly verifiable. Designed a consent architecture in which a permissioned ledger records patient grant, refusal and revocation as immutable events while all clinical data stays off-chain and encrypted per patient — no PHI on the ledger by design. Filed a proposed USCDI v7 data element with ONC on verifiable patient identity. Translates the Federal Register, HHS/CMS guidance and HITRUST requirements into architecture, controls and audit evidence.

Selected engineering scale

  • Self-managed estate of 67 public hostnames, 51 serving, on one AWS account.
  • FHIR R4 API with a SMART App Launch authorization server; patient and provider portals; an AI service layer; Hyperledger Fabric peer, orderer and CA; IPFS; a 337-table SQLite clinical store.
  • Two permissioned distributed ledgers in production, one operating a consent registry that records grant, refusal and revocation with no PHI on chain.
  • X12 005010 pipelines across eight transaction sets.
  • Designed, built, deployed and operated end to end by one person.

Credentials and standards work

  • Inferno (g)(10) self-test PASS, 317 of 317 subtests, April 2026 — the ONC Standardized API test suite. Standards validated: US Core 6.1.0, USCDI v3, SMART App Launch 2.0.0, Bulk Data 2.0.0. Public suite URL.
  • USCDI v7 public comment submission, April 2026 — proposed Patient Digital Identity Credential under the Provenance data class, listed publicly on the ONC Interoperability Standards Platform.
  • SAM.gov UEI MVKQW9DKKGB8 · CAGE 20KM5 · NPI 1194645564. AWS Business Associate Addendum active 18 August 2026.

Technical environment

AWS us-east-1 (compute, storage, messaging, SMS, email, voice) · Linux and systemd hardening · nginx · Fabric and EVM ledgers · FHIR R4 · X12 005010 · HL7 · SQLite · self-hosted mail, drive, search and monitoring · Git-based change control and CI release gates.

TRINH — Charles Trinh, technology architect
TRINH

Charles Trinh

Fractional CTO/CIO and Technology Architect — Healthcare Data, Identity & Regulated Infrastructure · Las Vegas, NV

Las Vegas, Nevada · US citizen · Fractional and project engagements

(702) 559-3489 Email Charles Trinh