Models inside the workflow · never outside the audit trail
Charles Trinh — Healthcare AI | Las Vegas, NV · Las Vegas, Nevada · Henderson · Clark County, Nevada · North Las Vegas, Nevada

Healthcare AI is mostly a data-plumbing problem wearing a model's clothes. The model is the small part; the provenance, the evaluation and the operational envelope are the work.
What this is
An AI feature in a clinical or administrative workflow has three components that are usually conflated: the model, the data pipeline that feeds it, and the decision rule that acts on its output. Most failures attributed to the model are failures in one of the other two.
The first requirement is provenance. Every input to a model that influences a decision should be traceable to its source, with a timestamp and a version. Without that, an adverse outcome cannot be investigated, and a model that cannot be investigated cannot be defended.
The second is evaluation. A model evaluated on a held-out set from the same distribution tells you very little about performance in production, where the distribution shifts. Evaluation has to be continuous, with monitoring for drift and a defined threshold at which the model is retrained or withdrawn.
The third is the operational envelope. A model that is accurate on average and wrong in a specific subgroup is not safe to deploy without knowing which subgroup. Stratified evaluation is not optional in a regulated context, and it is the finding most often missing from an AI deployment.
The fourth is human factors. A model output presented as a recommendation is acted on differently from one presented as information. The interface is part of the safety case, and a well-calibrated model behind a badly designed interface is a hazard.
Finally, healthcare AI sits inside HIPAA, and that constrains where inference can happen. Sending PHI to a third-party inference endpoint requires a business associate agreement and a scope that covers it, and that is a design decision rather than a procurement afterthought.
There is a regulatory dimension that shapes the design. A model that influences a clinical decision may be a medical device, and the boundary between a device and a tool is defined by intended use rather than by accuracy. Writing the intended use statement before building the model is what determines whether the work is a regulated submission or an internal tool, and discovering that boundary late is expensive.
Data governance is the second constraint. Training data carries its own consent and its own retention obligations, and a model trained on data that could not be used for that purpose is a liability regardless of its performance. Provenance on the training set is as important as provenance on the inference input.
The third is monitoring in production. A model performance degrades as the population changes, and the degradation is usually silent. Monitoring the input distribution, the output distribution and the outcome where it is observable is what turns a deployment into a managed service rather than a one-time launch.
Finally, the human factors question is not cosmetic. Whether an output is presented as a recommendation, a flag or a piece of information changes how it is acted on, and the difference is measurable. Designing the interface as part of the safety case, rather than as a layer applied afterwards, is what makes the deployment defensible.
A further consideration is the difference between a model that assists a decision and a model that makes one. An assisting model can be wrong and the clinician catches it; a model that makes the decision has no such backstop. The intended use statement has to say which it is, and the evaluation, the interface and the monitoring all follow from that answer rather than from the model architecture.
The final point is about cost. Inference has a per-request cost and a data egress cost, and a model that is accurate but uneconomic is not deployable at scale. Modelling the cost per decision before the build, rather than after the pilot, is what prevents a successful pilot from becoming an unaffordable service.
How I do it
I begin by mapping the decision the model is meant to support, who makes it, and what changes if the model is wrong. That framing determines the evaluation criteria far more than the model choice does.
From there I build the data pipeline with provenance as a first-class property: every input carries its source, its timestamp and its version, and the model's output carries a reference to the inputs it used.
Evaluation is designed as a continuous process with a held-out set, a production monitor and a drift threshold. Stratified results are reported by the subgroups that matter clinically, not only in aggregate.
The operational envelope is documented: what the model is for, what it is not for, what its known failure modes are, and what the escalation path is when it is wrong.
Finally, the deployment is reviewed against the data protection obligations, with the inference location, the agreement scope and the retention of inputs and outputs all stated explicitly.
The intended use statement is written first, because it determines whether the work is a regulated submission and what evidence is required. That single document shapes the rest of the programme.
Training data is reviewed for consent and retention before it is used, with the provenance recorded. A model trained on data that could not be used for that purpose is not deployed, regardless of performance.
Production monitoring covers input distribution, output distribution and outcome where observable, with alerts on drift rather than on accuracy alone, because accuracy is usually measured too late.
The interface is designed as part of the safety case: how the output is presented, what the user is expected to do with it, and what happens when the user disagrees. Those decisions are documented alongside the model.
Where the model assists rather than decides, the interface makes the basis of the output visible so the clinician can disagree with a reason. Where the model decides, the escalation path and the override are designed explicitly and tested.
The handover includes the intended use statement, the evaluation report with stratified results, the monitoring configuration and the operational envelope. Those four artefacts are what an auditor or a clinical safety officer will ask for, and having them ready is the difference between a deployment that can be defended and one that cannot.
Artifacts you keep
- A decision map naming the decision, the decision-maker and the consequence of error
- A data pipeline with provenance on every input and output
- Continuous evaluation with drift monitoring and a retraining threshold
- Stratified results by clinically relevant subgroup
- A documented operational envelope with known failure modes
Proof
- USCDI v7 proposal — Patient Digital Identity Credential — listed at isp.healthit.gov
- ONC (g)(10) self-test PASS, 317 of 317 subtests, April 2026 — public suite URL
- 104 clearinghouse enrolments across 53 jurisdictions
- 67 public hostnames, 51 serving, one AWS account
- SAM.gov UEI MVKQW9DKKGB8 · CAGE 20KM5 · NPI 1194645564
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
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
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
- Phone: 702-559-3489
- Email: charles@trinh.io
- Based in Las Vegas, Nevada · working remotely across the United States
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 — healthcare ai — the scope, the failure modes it prevents and how the work is measured. It sits in the X12 and EDI 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.
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.