Everything is real and traceable
Six things this build does that a chat wrapper cannot.
De-identified on device
OpenMed scrubs identifiers in your browser. The raw note never leaves it.
Trained on real notes
A note generator trained on open clinical corpora replaces the old hand-written templates.
Follow a fact anywhere
One code or dollar, shown as each of the four readers sees it, with full provenance.
Cross-portal handoffs
Flag a prior auth in the clinic; it lands in Revenue’s queue and ripples straight back.
Feedback flywheel
Every human correction becomes labeled training data, exportable as JSON.
Audited, not vibes
One verifier and constitution critic checks every report before it ships.
From messy note to auditable record
Intake + de-identification
Identifiers are scrubbed on your device before anything is sent.
Biomedical NER
OpenMed models tag diagnoses, medications, and labs with real confidences.
ICD-10-CM coding
A trained retriever and reranker link each mention to its billable code.
Risk + readiness
CMS readmission rates and a sourced claim readiness scrub, no fabricated denial score.
Generation + audit
Four portal writers, then one combined verifier and constitution critic before anything ships.
Patient: [redacted on device] · DOB: [redacted]
Encounter: 2026-07-01, follow up visit
HPI: pt presents with persistent chest pain radiating to left arm x 3 days.
PMH: Type 2 diabetes, hypertension, prior MI 2019.
Meds: Metformin 1000mg, Lisinopril 20mg, Aspirin 81mg
Codes: I20.9 Angina pectoris E11.9 T2DM I10 HTN
CPT 99214 CPT 93000 EKG
Risk: HRRP readmission rate 0.23
Claim readiness: PASS
▸ Patient summary generated
▸ Prior auth packet ready
▸ Claim submitted to payer
▸ Constitution audit: PASSED
Four portals, one shared encounter
The same coded record, rewritten for each reader. An action in one portal ripples through the others.
Patient
Plain-language explanations, medication guidance, and a personalized out-of-pocket estimate for the reader plan.
Physician
Suggested coding with sequencing, documentation prompts, prior authorization needs, and the readiness checklist.
Hospital
Claim construction with CMS amounts, review lane, expected reimbursement, and HRRP exposure.
Employer
Aggregated, anonymized population category, cost exposure, and benefit design, with no identifying detail.
One fact, four lenses
Click any code, dollar, or check in the demo and see the same truth as each reader sees it, with its provenance.
A plain-language card: what type 2 diabetes means and how to manage it.
Billable code E11.9 with provenance: linked from the note phrase.
Establishes the medical necessity for the billed services.
Rolls into the diabetes CCSR cohort, anonymized and aggregated.
Measured on open data
41%
exact ICD-10-CM code ranked first, on CodiEsp gold mentions (N = 3,615)
49%
correct code within the top five
0.44
mean reciprocal rank of the trained coder
Guardrails, not vibes
Codes cannot be invented
Every code comes from the official ICD-10-CM index and is revalidated against the CMS tabular before it appears.
Writers are audited
One combined verifier and constitution critic checks each report against the extracted facts before it is shown.
It abstains when unsure
Below a confidence threshold the pipeline escalates to a human coder instead of auto routing.