DRG Clinical Validation · Live protected demo

DRG validation review,
built as a workflow with the nurse reviewer in the loop.

A synthetic-data clinical validation assistant that helps a nurse reviewer work through DRG documentation cases. I built this from nearly 13 years as an ICU registered nurse combined with hands-on AI-enabled workflow systems experience. Runs as a protected demo behind Cloudflare Access — no real PHI, reviewer-in-the-loop, no automated overrides.

Problem

DRG validation review is documentation-heavy, exception-heavy, and easy to get wrong under time pressure. Reviewers spend a large share of their time reading unstructured clinical narrative, cross-checking documentation against criteria, and manually tracking what has and has not been checked on a given case.

Tooling in this space tends to either (a) drop AI suggestions into a freeform interface with no workflow, or (b) automate hard decisions that should belong to a qualified reviewer. Both modes erode trust and create rework. The opportunity is the middle: structure the workflow, capture evidence as the reviewer moves through it, and keep the human as final authority.

What was built

A synthetic-data clinical validation assistant designed around the way a nurse reviewer actually works the case: surface the relevant documentation, identify gaps against criteria, flag low-confidence inferences for human attention, and route the final disposition through a reviewer-controlled gate.

Every AI-assisted suggestion is treated as draft support content. The reviewer accepts, edits, or rejects it. The reviewer’s judgment is what advances state. The system never auto-applies a finding and never closes a case without a reviewer action.

Core workflow

  1. 1Synthetic case intake — narrative and structured fields ingested for review.
  2. 2Extraction — relevant clinical facts pulled into a structured shape.
  3. 3Normalization — terms and values standardised against a controlled vocabulary.
  4. 4Validation — automated checks flag gaps, conflicts, and low-confidence findings.
  5. 5Human review — nurse reviewer adjudicates each finding; nothing closes without an explicit action.
  6. 6Auditable disposition — final state, reasoning, and evidence are captured for the record.
Current status

Live protected demo — behind Cloudflare Access

The demo runs at drg-demo.twosevens.ai and is gated by Cloudflare Access. Visiting the URL directly will not grant access; reviewers must request approval first. Use the gateway below — that is where access requests are accepted.

Safety boundary

The boundaries below are not aspirational — they are how the demo is actually run.

  • Synthetic cases only

    No real patient records, no protected health information, no de-identified production data. Cases are constructed for demonstration.

  • Reviewer-in-the-loop, always

    AI output is draft support content. A qualified nurse reviewer is the final authority on every disposition.

  • Not approved for real PHI

    This is a demonstration surface. It is not a HIPAA-cleared production system and makes no PHI-readiness claim.

  • Access by approval only

    Exact-email allow-list managed by hand. External reviewers are added only after a request is submitted and reviewed.

Architecture

The shape, not the secrets. No internal hostnames, no network-level identifiers, no Cloudflare Access policy IDs.

Web application

Next.js application that renders the reviewer surface and serves the validation workflow. Containerised for reproducible build and deploy.

Persistence

Postgres holds synthetic case content, validation state, and reviewer activity. Schema is purpose-built for demo cases — no PHI columns, no real patient identifiers.

Access edge

Cloudflare Access fronts the demo with an exact-email allow-list and a tunnel route. Approval is manual; nobody reaches the application surface without prior review.

Workflow gating

Extraction → normalization → validation → human review. Each transition is explicit. The nurse reviewer is the final authority before any output is treated as decision-ready.

CI and operational discipline

Lint, type-check, build, and audit gates run on every change. Releases are reviewed; nothing ships ad-hoc.

Access model

External reviewers only get access after an explicit approval. There is no public path to the live demo.

Exact-email allow-list

External reviewers receive access only after a request is submitted through the gateway and the email is added to the allow-list by hand.

Manual approval window

Requests are typically reviewed within 24 hours. There is no self-service signup and no automated provisioning to the live demo.

Scope-limited session

Approved reviewers see the synthetic-data demo only. Operator surfaces, configuration, and any internal tooling stay out of scope.

Disclaimers — read these

  • Not for use with real protected health information. The hosted demo accepts synthetic clinical scenarios only.
  • Synthetic / demo data only. Cases shown to reviewers are constructed examples, not derived from real patient records.
  • Nurse reviewer remains the final authority. Any AI-assisted output is draft support content, not a coding or clinical decision.
  • Output is draft support content. It exists to accelerate review, not to replace it.

Request protected demo access

Reviewers, healthcare AI evaluators, and hiring teams: submit a request through the gateway. Access is granted by hand after a manual review of the request.

Request protected demo access

Background — Why I can build this responsibly

Why this work is grounded in real clinical operations

I built this work from nearly 13 years as an ICU registered nurse, including about 11 years in travel nursing across multiple hospital systems. That background gives me a practical view of how clinical documentation, handoffs, staffing pressure, fragmented workflows, and operational risk actually show up on the floor.

I combine that clinical experience with hands-on work in AI-enabled workflow systems, distributed infrastructure, automation, and production-style software design. My focus is not replacing clinical judgment. It is building tools that reduce friction, preserve accountability, and keep humans in control of high-consequence decisions.

What I help teams do

  • Diagnose where workflows break, stall, or create rework
  • Redesign operational processes into cleaner states, rules, and checkpoints
  • Implement AI-assisted workflows with validation and human oversight
  • Test workflows for edge cases, exception paths, and failure conditions
  • Improve operational consistency without hiding risk
  • Support go-live, adoption, and cross-functional handoffs
  • Translate operational pain into implementation-ready systems and rollout plans

Related proof

The same workflow discipline applied to adjacent problems — extraction-to-structured-output automation, workflow gating with evidence capture, and reliability practice on distributed operations.

AI Autolister

Inconsistent source inputs become validated, structured outputs through extraction → normalization → validation → human review.

What the buyer gets

Manual processing burden goes downAutomation handles repetitive extraction and normalization
Output consistency goes upRule-based normalization plus AI-assisted enrichment with validation thresholds
Quality control is preservedHuman-review gate sits between automated output and downstream publication

Workflow-Gated Operations

Workflow-heavy operations get staged state, required-action gates, evidence capture, and controlled closeout.

What the buyer gets

Process drift goes downRequired-action gates make step ordering explicit, not tribal
Audit trail gets cleanerEvidence capture is attached to state transitions, not retro-built
Rework from missed steps dropsCloseout is blocked until validation checks pass

Reliability & Pipeline Operations

A distributed environment operated as production: service separation, recoverability, queue-backed execution.

What the buyer gets

Failures are recoverable, not catastrophicBatch checkpointing, fault tolerance, queue-backed orchestration
Workloads scale without rewritingGPU hosts, application services, and data layers separated cleanly
Operations look like a productMonitoring, observability, and reliability treated as first-class concerns

Capabilities

What I bring to employer teams working on clinical AI, workflow redesign, and infrastructure-heavy automation.

Workflow & implementation

Clinical workflow optimizationProcess redesignRequirements gatheringWorkflow testing & validationFailure-point analysisException-path designChange managementGo-live executionUAT supportSOP & playbook design

Healthcare AI delivery

Healthcare AI implementationHuman-in-the-loop AIStructured output generationReview-gate designUnstructured-to-structured transformationOCR / extraction workflowsData normalization & validation

Technical systems

Distributed systems operationsQueue-backed processingBatch & orchestrationReliability & observabilityAPI-oriented systems thinkingFault tolerance & recoverability
Available for evaluation

Work with me on clinical AI workflows

I take on workflow audits, clinical AI implementation, and technical solutions engineering through scoped work with concrete artifacts.