Identified the physician workflow, secondary users and the brief → journey → evidence → question path before choosing the technology.
← All projects
Healthcare AI · Enterprise architecture
ClinIQ Nexus
Clinical intelligence on a governed lakehouse: what happened, what matters, and where the evidence is.
At a glance
A physician-first workflow, designed end to end.
My roleProduct strategy, solution architecture, hands-on engineering and release governance
Primary userPhysicians reviewing charts before follow-up or after discharge
PlatformDatabricks lakehouse, FastAPI and React / TypeScript
Delivery disciplineValidated release with staged acceptance and a rehearsed rollback
The problem
The patient story is fragmented.
A physician preparing for a follow-up visit or reviewing a patient after discharge has to reconstruct what happened across encounters, notes, laboratory results, medication changes, portal conversations, clinical PDFs and imaging reports.
Tables leave the synthesis to the clinician. Untraceable AI creates a different risk: a fluent answer that cannot be checked against the record—or that turns missing information into a negative finding.
The solution
Context with evidence attached.
ClinIQ Nexus brings structured and unstructured clinical sources into one patient-scoped workspace: a physician brief, longitudinal journey, trends, medication history, communications, document evidence and an assistant that cites its sources.
The governing idea
Use deterministic answers before generative ones. Narrative AI claims must be cited, validated and independently verified—or the interface shows the underlying evidence instead.
My role
How I led it.
I treated this as a product and architecture exercise—not simply an AI demo—so the workflow, safety boundaries, governance and release discipline were designed together.
No diagnosis, treatment, dosing or image interpretation; read-only access; synthetic labelling; “not documented” instead of guessing.
Replaced placeholder-quality data with a coherent synthetic dataset whose notes, PDFs and structured facts agree.
Moved through discovery, design, implementation, evaluation, candidate deployment, acceptance and promotion.
Worked within the available platform, documented deferred capabilities, and clearly separated implemented features from proposed work.
Combined automated checks with human review of generated claims and a rehearsed rollback to the prior version.
Physician workflow
From signal to source.
The interface is organized around the questions a physician asks: what needs attention, what changed over time, and what evidence supports the answer?
PDF intelligence
Evidence you can open.
- Extracts text page by page and retains page provenance.
- Verifies patient and encounter identity on every page.
- Quarantines a mismatch before it reaches any patient view or AI index.
- Flags image-only scans as not machine-readable; no OCR is implied.
- Opens the original PDF at the cited page.
Grounded AI
An answer has to earn its place.
Policy and structured queries run before generation. Narrative answers are patient-scoped, claim-based and fail closed if evidence does not support them.
Cross-patient, diagnostic, treatment, dosing, prognosis and image-interpretation requests are handled before any model call.
Questions about medications, labs, messages, follow-up and imaging inventory are answered directly from records.
Lexical and semantic ranking runs only over the selected patient’s records, with dates steering “recent” and “initial” questions.
The model returns individual documented claims with citations—or marks information as not documented.
Citations, numbers, dates, lexical support, safety language and completeness are validated, then a second model checks support.
If a retry cannot pass the same checks, the user sees the relevant evidence and the reason no answer was shown.
Architecture
Governed ingestion, patient-bound access and verified answers.
Synthetic sources land in a Unity Catalog volume. A serverless ingestion notebook verifies PDFs and DICOM headers, creates page-aware retrieval chunks and publishes Delta tables only when all 14 integrity checks pass. The application reads through a SQL warehouse with patient-bound parameters.
| Architecture decision | Why it matters |
|---|---|
| One Databricks App with React and FastAPI | A persistent physician workflow, evidence drawer and timeline interactions in one deployable unit and one application identity. |
| Deterministic first, generative second | Structured questions map directly to governed records; generation is reserved for narrative synthesis. |
| Second-model verification plus grounding checks | A citation identifier alone is not proof that a source supports a claim. |
| Quarantine misfiled PDFs at ingestion | Patient-isolation failures are stopped before publication, rather than hidden at the interface. |
| Read-only, least-privilege application access | The application cannot modify clinical data and only accesses the required schema, volume, warehouse and model endpoints. |
Validation
Demonstrated, not claimed.
These figures are engineering evidence for the synthetic implementation—not clinical-effectiveness, productivity, adoption or ROI claims.
Application and pipeline checks passed.
Cross-source clinical facts agree.
End-to-end checks passed against the deployed application.
Structured, narrative, missing-information and policy cases.
What the AI evaluation showed
Every cited source came from the selected patient, no citation failed to resolve, all 18 narrative answers passed verification, all 6 questions about absent information returned “not documented,” all 9 advice, image-interpretation and cross-patient requests followed policy, and no imaging finding was invented.
Walkthrough
See the workflow in 90 seconds.
The captioned walkthrough was captured from the deployed Databricks application. It contains no audio and uses synthetic records throughout.
Next in a real clinical program
Production readiness would be a separate phase.
- On-behalf-of-user authorization with row-level security tied to clinical roles.
- OCR evaluated against ground truth before scanned documents become citable.
- MLflow tracing and per-release evaluation tracking.
- Managed Vector Search when the corpus outgrows in-app ranking.
- A licensed, de-identified imaging library kept separate from synthetic patients.
- Formal clinical-safety review and usability testing with physicians.
Important disclosure
ClinIQ Nexus is a technology demonstration built on synthetic data. It is not a production clinical system, an EHR integration, a diagnostic tool or HIPAA-certified software, and it makes no clinical-outcome or financial claims.