AI

Charles Trinh — Clinical Documentation AI | Las Vegas, NV · Las Vegas, Nevada · Henderson · Clark County, Nevada · North Las Vegas, Nevada

A note draft panel showing supplied facts, generated sentences and clinician edits
Charles Trinh — Clinical Documentation AI

Drafting a note is easy. Making the draft provably faithful to what was said, attributable to the clinician who signed it, and reviewable years later is the engineering. I design the whole chain, not just the prompt.

Where a note draft goes wrong

Documentation models fail in ways that reading the output almost never reveals. They confidently state a dose that was never mentioned, fold two encounters into one narrative, or carry an earlier diagnosis forward because it was present in the context window. The sentence reads well and is wrong. I treat every generated sentence as a claim that must be traceable to a supplied fact, and I measure the rate at which drafting produces claims with no such fact behind them on a reviewed corpus, per note type.

Carry-forward is the most damaging of these because it repeats silently across thousands of notes. A patient with a resolved condition that keeps reappearing as active looks like a coding problem to the finance team and a clinical one to the treating team, and both are right. The fix is structural: the context supplied to the model names what is active, what is historical, and what it does not know, and I test explicitly that a resolved item does not reappear in a draft when the supplied facts mark it resolved.

The failure mode I often find in existing deployments is a prompt that says not to invent facts, with no test behind it. Instructions in a system message reduce the rate but do not eliminate it, and nobody can tell you how much they reduce it on your notes. That number is what I produce first, on your templates, your specialties and your note length distribution, before any change to the pipeline.

The supplied-fact model

A clinical-documentation-ai pipeline should be split at a clear seam: an assembler that collects the facts for one encounter, and a drafter that turns those facts into prose. The assembler pulls from the encounter record, structured observations, medication lists, prior notes and any consented audio transcript. Each fact carries its source, its timestamp and a stability tag. The drafter receives only the assembled facts and never the raw store, which keeps the model's inputs enumerable and replayable.

Because the input is an explicit fact set, the output can be checked against it. A second pass extracts the claims from the draft and asserts each one against the set. Claims that fail are dropped from the draft rather than flagged for a clinician to catch, because the clinician will not catch them. On a review of ambient-scribe-evaluation runs I usually find the correction rate for unsupported claims is lower when the checker runs than when a prompt instruction asks the model to police itself.

The seam also makes the system testable: an encounter can be replayed through the assembler, the drafter swapped independently, and the output compared against a fixed input. Without that boundary a change to either half invalidates every earlier evaluation, and teams stop evaluating once the baseline keeps moving under them. It is also the place where a vendor swap is contained, because only the drafter changes and the fact set becomes the fixed comparison point for the new supplier.

Edit capture and the clinician's signature

The draft is not the note. The clinician edits it, signs it, and that signature is the legal act. I capture every keystroke-level change as a diff with a timestamp, so the record can show what the model proposed and what the clinician accepted. That record is the evidence that the clinician, not the model, authored the note, and it is what a malpractice defence, a coding audit or an internal review will ask for first.

Edit capture also produces the most useful evaluation signal you will ever have: if clinicians routinely delete or rewrite a particular block, that block is wrong for your service line regardless of what any offline score said. I aggregate edits by block and by specialty on a weekly cadence and feed the result back into the assembler and the prompt. Declining edit volume on a stable note type is one of the few signs that a documentation tool is genuinely getting better.

Approval buttons are where documentation tools go soft. A button labelled accept that sends an unread draft to signature is a defect, and the same applies to a one-click accept that fires before the draft has loaded. I test the sign flow against a real rendered interface, not the handler, and I require that an unedited draft cannot be signed without a separate confirmation that names it as unedited.

Measuring faithfulness, not fluency

The measurement I run first is omission and fabrication on a stratified sample. For each note type, a clinician reviewer reads the supplied facts and the draft side by side and records unsupported claims, missing facts that change the clinical picture, and wrong attributions. I report the three rates separately because they move independently: a model prompted to be cautious lowers fabrications and raises omissions, and teams read the single blended score as progress.

Attribution accuracy matters clinically. A draft that merges the patient's report and the clinician's assessment into one voice can carry the patient's speculation into the chart as the clinician's conclusion. I treat this as a fabrication class of its own and measure it on notes that contain both. The repair is structural tagging in the supplied fact set, and it is testable in exactly the same harness, so it does not need a new evaluation programme.

Fluent drafts dominate reviews. Reviewers skim, and a note that reads naturally passes review more often than a clumsy but accurate one. I offset this by hiding fluency cues from at least one review pass: the reviewer sees the facts and the extracted claims, not the prose, so the judgement is on claims rather than style. Running both passes on the same notes changes the unsupported-claim rate materially and tells you who is reviewing what.

Audit trail for the first external visit

The first external visit is usually an auditor with a sample of notes and one question: where did this sentence come from. A supplied-fact system answers that in seconds. Each draft stores its fact set, drafter version, prompt version and checker output; each signed note stores the draft-to-note diff; and the whole record is immutable and readable by someone who is not a developer.

That same record answers a HIPAA audit under hipaa-privacy-security-gap. The audit trail shows who accessed the encounter, which model and version processed it, where inference ran, and whether the model vendor met the terms required for a business associate. If any of those cannot be retrieved for a note older than a year, the design has a gap that is not fixable by adding a log line later. I test retrieval against a real twelve-month-age record.

The artefact handed over at the end is a note-level provenance report: for a sample of notes, the fact set, the draft, the diff, the signature and the retrieval time for each. It is usually enough to close the first visit, and it becomes the fixture that catches a silent breakage when a vendor ships a new model and the trail stops recording something.

Change control when the model vendor ships

Documentation models are hosted services and they change without asking you. A model update can shift the style, the length, the willingness to invent, and the rate at which the checker's own extractor recognises claims. None of these show up in your dashboards because nothing is watching them. I maintain a frozen reference corpus of encounters and a fixed set of expected facts, and I re-run evaluation against both whenever the drafter or the checker version changes, whether or not anyone told me it did.

Pinning helps but does not solve it. A pinned model gets deprecated, and the deprecation date is rarely the one your contract assumed. I build the pipeline so that changing the drafter is a configuration change with a new version number, and so that a failed evaluation blocks the release rather than generating a warning. The same conformance gate that blocks a code release blocks a model swap, and it is the reason a vendor announcement becomes a scheduled task rather than a fire.

This work sits inside the programme described in ai-model-governance-healthcare and uses the review pattern from human-in-the-loop-review. Where a documentation vendor offers nothing beyond a demo, ai-vendor-diligence is the check on what it records and retains about your encounters. That check should run before the first draft reaches a clinician, not after the pilot has been extended for a third month.

What the engagement leaves behind

At the end of an engagement you have the assembler, the drafter wrapper, the claim checker, the evaluation harness and the provenance report, all in your repository and run from your own pipeline. None of it depends on a dashboard I own. The evaluation harness is the piece I care about most, because it is what lets your team answer the next vendor question without me, and it is what I would expect to keep being useful after the engagement ends.

My background is building and operating backend systems, not practising medicine. I will not tell you whether a draft is clinically correct; I will build the process by which your clinicians decide that, and the record that proves it happened. That division is honest and it is also the only useful one, because a builder replacing clinician judgement would be the failure mode the whole design exists to prevent.

Engagements start with a measured read of your notes and a cost estimate for adding the checker, not with a demo. The estimate covers the evaluation corpus work, which is usually the largest single item and the one most often skipped. If you have a live documentation tool, the first deliverable is the unsupported-claim rate against your own notes, measured, because most teams have never seen that number.

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 — clinical documentation ai — the scope, the failure modes it prevents and how the work is measured. It sits in the Security and Compliance 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