Zahabiya Marvi — Senior Product Manager

Portfolio

Selected work I've shipped

↑

ReferralFlow AI

Healthcare AI · Intake automation

ReferralFlow AI — Referral to intake automation

Intake coordinators upload referral forms — faxes, PDFs, scanned documents — and an LLM extracts clinical and patient data, auto-populating the electronic intake form with per-field confidence scoring. Pilot complete; currently refining edge cases like handwritten forms before full production rollout.

91% extraction accuracy 87% effort reduction 15 min → <2 min per referral

OCR + LLMConfidence scoringHuman-in-the-loopEnterprise SaaS

View case study →

♪

NoteFlow AI

Healthcare AI · Clinical documentation

NoteFlow AI — Clinical note generation

Wound care clinicians speak or type their assessment in natural language. NoteFlow generates a structured, HIPAA-compliant clinical progress note with all 9 required fields — ready to paste into any EMR in under 2 minutes. Live in production; nurses typically use it after the visit, dictating from their car before driving to the next patient.

67% documentation time reduction 21%→8% claim denials 45→15 min per visit

Voice-to-noteLLM structuringHIPAA-compliantHuman review

View case study →

Concepts & explorations

Things I've designed and prototyped

Product concepts I've developed independently — not shipped at work, but built to think through real problems and sharpen how I approach AI product design.

→

ClearPath

Product concept · Interview case study

ClearPath — Discharge coordination

A product concept I developed for a healthcare company interview panel — an AI-powered discharge coordination platform with FHIR R4 integration, real-time care team dashboard, AI-generated blocker summaries with source citations, and ranked recommended actions. Built a fully interactive prototype to bring it to life.

60–70% coordinator time on status-chasing 2–4 hrs avg delay per patient

FHIR R4React prototypeDischarge planning

View case study →

Insights Engine

In development

Customer Insights Engine

An AI tool for surfacing product pain points and customer signals from Reddit conversations. Currently in development — check back soon.

n8nClaude APIReactComing soon

About

A little bit about me

I'm a people-first PM — which in practice means I spend a lot of time understanding how people actually experience their work, not just what they say they need. That instinct for the human side of a problem has shaped most of the products I've built.

Most of my career has been in healthcare and SaaS. It's a good fit — the problems are real, the stakes are high, and there's rarely a shortage of processes that have been broken for years and quietly accepted as normal.

At United WoundCare Institute I owned product strategy for an enterprise wound care platform supporting $200M+ ARR. The work was operationally deep — referral intake running on faxes, clinical documentation taking nurses 30+ minutes per visit, wound imaging done by hand. We built AI tools for each of these and moved the numbers meaningfully. Before that, I was at Intuit building analytics tooling for security teams, and earlier I ran product at an e-commerce company where a focus on customer research and experimentation drove 224% revenue growth year over year.

I have an MBA from Berkeley Haas and I care about getting the business case right alongside the user experience — not as an afterthought.

What I tend to gravitate toward:

AI tools that make real work meaningfully easier

Workflows that are manual and fragmented and have been that way too long

Analytics products that help people make better decisions faster

0→1 builds where the problem definition is still evolving

Healthcare — because the consequences of getting it wrong are real

← Back to portfolio

Healthcare AI · Intake Automation · UWI

ReferralFlow AI

AI pipeline that extracts 20 structured fields from faxed referral PDFs and auto-populates the electronic intake form with per-field confidence scoring. Pilot complete — currently in refinement, working through edge cases like handwritten forms before full production rollout.

87%

overall extraction accuracy

91%

accuracy on NPI & diagnosis

85%+

intake time reduction

20 fields

extracted per referral

→ View prototype (Referral-Intake tab)

Problem

Solution

Try it

Results & testing

Business case

Deployment

The problem

Referral intake is still a manual, fax-driven, error-prone process

~15 min

to manually process a single referral — extracting and re-entering data from faxed PDF into the EMR

Netsmart / industry

18–21%

of claims denied at UWI, with intake data errors as a primary driver

Internal UWI data

Dozens/day

of referrals processed per coordinator, each requiring field-by-field manual entry

UWI ops

"How might we eliminate manual data entry from the referral intake process — giving coordinators a pre-populated intake form with field-level confidence signals, so they can verify rather than transcribe?"

Each referral packet arrives as a faxed PDF. An intake coordinator extracts demographics, insurance details, clinical information, and provider data — then re-enters all of it into the EMR one field at a time. Across dozens of referrals per day, this is the dominant source of intake coordinator time, and errors at this stage directly propagate to claim denials downstream. The coordinator isn't slow. The workflow is broken.

The solution

AI-powered extraction with per-field confidence scoring

📄

PDF text extraction — not OCR

Referral PDFs are processed using pdftotext, extracting digitally embedded text directly. OCR is not used — it adds noise and error. pdftotext works on the vast majority of digitally generated fax referrals. Handwritten or scanned-only forms are a documented roadmap item.

🧠

Claude-powered structured extraction

Extracted text is sent to Claude with a structured prompt to identify all 20 intake fields. Claude understands clinical context — it correctly parses wound type, diagnosis codes, and visit frequency from free-text referral language, not just from labeled form fields.

⚖️

Weighted confidence scoring algorithm

Each field carries a confidence score calibrated against a weighted algorithm where high-stakes fields (NPI, Medicare ID, primary diagnosis) carry more weight toward the overall confidence badge. The PM owns the field weights — they reflect clinical and billing importance, not just extraction difficulty.

🚦

Three-state confidence UI

Each field renders in one of three states so coordinators know exactly where to focus review time.

✓ Green — Confirmed

NPI: 1234567890

Confidence ≥90%. Extracted cleanly — accept as-is.

⚠ Yellow — Review

ZIP: 9284_

Confidence 80–89%. Extracted but incomplete — flag for coordinator review.

✗ Red — Missing

Medicare ID: Not found

Field not present. Coordinator must enter manually.

Try it

Experience the prototype

Live interactive prototype

ReferralFlow AI — Referral to Intake Automation

Upload a referral PDF and watch the intake form auto-populate with color-coded confidence indicators. The demo includes a test referral designed to show all three confidence states — green, yellow, and red.

→ Open CareFlow AI prototype

Select the Referral-Intake tab after opening

ReferralFlow and NoteFlow currently share one prototype URL — standalone URLs coming soon

Results & testing

Current prototype performance

91%

Accuracy on NPI & primary diagnosis

87%

Overall accuracy across all 20 fields

85%+

Time reduction on typed forms

Referral test set with known ground truth

📋 Eval set design

30 referral PDFs with known correct values across all 20 fields

Mix of clearly typed, partially filled, and ambiguous forms

Accuracy thresholds defined before results were reviewed

Coordinator correction dataset built from flagged fields

📊 What we measure

Field-level extraction accuracy per field

Confidence calibration — 90% confidence = 90% correct?

False negative rate — fields returned null that had a value

Override rate by field — which fields coordinators correct most

Issues surfaced in testing

Issue

Handwritten and scanned-only forms returned empty extraction

pdftotext extracts digitally embedded text only. Fully handwritten referrals return blank output. Mitigation: documented as known gap; roadmap includes Claude Vision as a fallback with UI flag alerting coordinators.

Issue

Confidence calibration on date fields was inconsistent

Claude returned high confidence on dates formatted incorrectly. Mitigation: added post-extraction date normalization and a regex validation layer that independently verifies date format before rendering confidence.

Learning

New agency templates require prompt updates

Different referring agencies use different form layouts. Extraction accuracy on new templates starts around 75–80% and improves with prompt refinement. Informs the production rollout strategy: onboard agencies incrementally with a validation period.

Business case

The financial case for fixing intake

~15 min

Manual processing time per referral

<2 min

With ReferralFlow AI on typed forms

18–21%

UWI claim denial rate — intake errors a primary driver

$5,000

Avg revenue per patient episode at risk per denial

Coordinator efficiency

A coordinator processing 20 referrals/day spends ~5 hours on manual data entry. ReferralFlow reduces that to under 40 minutes — freeing the remainder for relationship management, follow-up, and exception handling.

Denial rate reduction

Intake errors — missing Medicare IDs, incorrect NPI numbers, incomplete prior auth — drive claim denials. A 1% improvement in denial rate across UWI's ~104,000 monthly visits represents ~$160K in recovered monthly revenue.

Engineering de-risking

The prototype validates AI feasibility before committing to a full production build — avoiding a costly engineering investment in a system that might require fundamental architectural changes post-launch.

Deployment plan

Prototype → pilot → production, gated on outcomes

Pilot

✓ Complete

30-referral eval set with known ground truth

Coordinator correction dataset built

91% accuracy on NPI + diagnosis confirmed

Gate met

Edge case refinement

In progress

Handwritten and scanned-only form handling

New agency template tuning

Date field confidence calibration fixes

Gate: stable accuracy across form types

Production build

Weeks 13–24

EMR write-back with idempotency controls

Claude Vision fallback for handwritten forms

Expand to all agencies with per-agency tuning

Measure: denial rate delta, coordinator NPS

Confidence calibration drift

Claude model updates may shift calibration between predicted confidence and actual accuracy. Mitigation: monthly evaluation set run in production to detect drift.

EMR write-back integrity

Writing extracted data directly to EMR without review creates patient safety risk. Mitigation: human review permanent and mandatory before any write.

HIPAA / BAA requirements

Claude API processes PHI from referral documents. A BAA with Anthropic is required before production — negotiation must begin during pilot.

← Back to portfolio

Healthcare AI · Clinical Documentation · UWI

NoteFlow AI

Wound care clinicians speak or type an assessment. NoteFlow generates a fully structured, HIPAA-compliant clinical progress note — all 9 required fields, ready to paste into any EMR in under 2 minutes. Live in production; nurses typically dictate from their car after leaving a patient's home.

67%

documentation time reduction

45→15 min

per visit

21%→8%

claim denial rate

9 fields

structured every note

→ View prototype (Clinical-Note tab)

Problem

Solution

Try it

Results & testing

Business case

Deployment

The problem

Wound care documentation is time-consuming, error-prone, and driving clinician burnout

~45 min

Documentation time per wound care home visit — major component of the ~61 min average visit

Swift Medical / industry

27%

Annual RN turnover in home health — documentation burden a consistently cited driver

Home Health Care News 2024

21%

UWI claim denial rate — incomplete or incorrectly structured notes a key driver

Internal UWI data

After each home visit, wound care nurses sit in their car and document before driving to the next patient. It's actually one of the quieter moments in their day — no background noise, no distractions — but the documentation itself is still a significant cognitive load. The note has to capture 9 required clinical fields in exactly the right format. Miss one, and you risk a claim denial. Get the wound stage wrong, and you risk a full episode audit where the payer reviews every visit note and can claw back up to $5,000 in reimbursement.

"How might we allow wound care clinicians to document an assessment through natural speech or text — and receive a fully structured, compliance-ready clinical note in seconds, with no cleanup required?"

Why existing dictation tools don't solve this

Capability

Dragon / generic dictation

NoteFlow AI

Output

Raw transcript — requires manual cleanup

Structured clinical note — ready to paste into EMR

Clinical understanding

None — transcribes words, not meaning

Understands wound staging, tissue types, compliance requirements

Required fields

No awareness — omissions pass through silently

All 9 fields always present in output

The bottom line

Dragon transcribes what you say

NoteFlow writes the note

The solution

Voice or text in → structured clinical note out

🎙️

Voice input

Clinician dictates

📝

Transcription

OpenAI Whisper

🧠

Clinical structuring

Claude API

📋

Note generation

9-field output

✅

Clinician review

Human-in-the-loop

Key distinction: Whisper handles transcription. Claude handles everything else — wound staging classification, tissue type interpretation, documentation structure, compliance formatting. Whisper is a transcription engine. Claude is the clinical reasoning layer that turns that transcript into a finished document.

🧠

Clinical context and medical knowledge

Claude understands wound staging (Stage 1–4, unstageable), tissue types (granulation, fibrinous slough, necrotic), and wound care terminology without domain-specific fine-tuning. It interprets clinical language and maps it to structured documentation fields correctly.

⚠️

Infection risk detection with alert

A secondary classification layer flags signs of wound infection (seropurulent exudate, periwound erythema) and triggers an alert to the supervising clinician. Designed for recall — better to flag a non-infected wound than miss an infected one.

The 9 required clinical fields

01Wound location & anatomical site

02Wound stage / classification

03Wound dimensions (L × W × D cm)

04Wound bed tissue type

05Exudate type & amount

06Periwound skin condition

07Treatment applied

08Pain level (0–10 scale)

09Clinician assessment & plan

Try it

Experience the prototype

Try this wound assessment script

"Presenting with a Stage 3 sacral pressure injury measuring 4 by 3 centimeters, depth 1.2 centimeters. Wound bed with 60% granulation and 40% fibrinous slough. Moderate seropurulent exudate. Periwound maceration noted. Pain level 4 out of 10. Treated with sharp debridement, wound irrigated with normal saline, calcium alginate with silver applied. Plan to reassess in 48 hours."

Live interactive prototype

NoteFlow AI — Clinical Note Generation

Speak or type a wound assessment and watch a fully structured 9-field clinical note generate in real time. Try the wound care script above to see how NoteFlow handles complex clinical terminology.

→ Open CareFlow AI prototype

Select the Clinical-Note tab after opening

NoteFlow and ReferralFlow currently share one prototype URL — standalone URLs coming soon

Results & testing

Prototype performance and evaluation

67%

Documentation time reduction (45→15 min)

9/9

Required fields present in test outputs

21%→8%

Claim denial rate improvement

Hallucinated clinical details in test runs

📋 What we measure

Structured field coverage — all 9 fields present

Clinical accuracy — correct wound staging and terminology

Override rate by field — where clinicians correct most

Hallucination check — output contains no details not in input

🔬 How we evaluate

Automated field presence check across outputs

Wound care nurse scores sample outputs 1–5 on accuracy and usability

LLM-as-judge: Claude evaluates against a clinical documentation rubric

Side-by-side comparison with manually written notes

Business case

Documentation quality is a direct revenue lever

~$5,000

Avg revenue per patient episode at risk per documentation deficiency

~$2,300

Per 30-day billing period clawed back in a full episode audit

$500K+

Recovered monthly revenue from 1% denial improvement at UWI scale

27%

Annual RN turnover — documentation burden a cited driver

Clinician time savings

67% reduction in documentation time (45→15 min) across 1,000+ UWI clinicians translates to thousands of recovered hours per week — redirectable to patient care or additional visits.

Denial rate and audit risk reduction

Incomplete or incorrectly structured notes trigger denials and full episode audits — where payers review every visit note and can claw back an entire billing period ($2,300) or episode ($5,000). NoteFlow ensures all 9 required fields are always present.

Workforce retention

27% annual RN turnover — with documentation burden as a cited driver — represents significant recruitment and training cost. NoteFlow reduces post-visit documentation load, addressing one of the most cited contributors to burnout.

Deployment plan

Prototype → pilot → production, gated on outcomes

Prototype validation

✓ Complete

20-script eval set across wound types

Wound care nurse review and scoring

All 9 fields consistently present — confirmed

Gate met

Pilot

✓ Complete

Rolled out to a subset of clinicians

Override rate tracked by field

Adoption positive, trust signals strong

Gate met

Production — live

In use now

Nurses dictate post-visit, typically from their car

Continuous prompt refinement from override data

Claim denial rate improvement tracked ongoing

21%→8% denial rate improvement

Home environment audio quality

Background noise degrades Whisper transcription accuracy on clinical terms. Mitigation: text input path as fallback; Whisper prompt tuning with wound care terminology.

Hallucination in generated notes

Claude may generate plausible-sounding detail not in the input. Mitigation: permanent human review before EMR entry; instruction prompt returns "not documented" rather than inferring missing fields.

Model drift after updates

Claude model updates can shift output format or clinical interpretation. Mitigation: monthly evaluation set run in production; automated field coverage check alerts team when format changes.

← Back to portfolio

Product concept · Interview case study

ClearPath — AI Discharge Coordination

An AI agent that synthesizes EHR, nursing, PT/OT, pharmacy and case management data to surface the specific blocker preventing discharge — and the next action to unblock it.

60–70%

of coordinator shift on status-chasing

2–4 hrs

avg avoidable delay per patient

~$900

excess cost per delayed discharge

3 roles

Coordinator · patient · care team

→ View live prototype

Problem

Solution

Try it

Business case

Deployment

The problem

Discharge readiness lives across six systems. No one has the full picture.

60–70%

of a coordinator's shift spent on status-chasing, not care coordination

Advisory Board / NEJM Catalyst

2–4 hrs

avg avoidable discharge delay per patient — ~$900 excess bed-day cost per delay

NEJM Catalyst / Advisory Board

20–25%

annual coordinator turnover — cognitive burden and fragmentation a cited driver

ACMA / industry

"How might we give discharge coordinators a real-time, AI-synthesized view of each patient's discharge readiness — surfacing the specific blocker, its source, and the exact next action needed to unblock it?"

Discharge readiness data lives across six systems — EHR physician orders, nursing notes, PT/OT documentation, pharmacy, case management, and transport scheduling. No single system synthesizes across all of them. Existing tools like Qventus predict when a patient might leave but don't identify the specific blocker preventing discharge today. Epic/Cerner built-ins rely on structured fields only and can't read free-text nursing notes or case management documentation. The gap isn't better dashboards. It's synthesizing what dashboards can't read.

The solution

ClearPath — an AI discharge coordination agent

⊕

Real-time discharge coordination board

All patients on a unit in one view, filterable by status (Escalated / Blocked / Ready). Each patient shows discharge readiness %, wait time, and active blockers. Proactive alerts surface patients waiting 4+ hours before coordinators go looking.

🔍

AI blocker identification with source citation

ClearPath synthesizes Epic EHR (FHIR R4) and CareManager Pro (HL7 ADT) to identify the specific blocker — not a status flag. Every blocker includes AI confidence score, data source, note freshness, and responsible clinician.

→

Ranked recommended actions with responsible party

For each blocker, ClearPath generates prioritized next actions assigned to the right person — coordinator, case manager, bedside nurse, or pharmacist. The coordinator makes the final call — always.

💬

Patient-facing plain-language discharge update

A second view generates a plain-language message for the patient explaining their discharge status clearly. AI-generated, reviewed by coordinator before delivery — reducing patient anxiety and improving adherence.

Try it

Experience the prototype

Live interactive prototype

ClearPath — Discharge Coordination Board

Built and fully interactive. Click through the patient board, explore individual discharge statuses, active blockers with AI confidence scores, recommended actions, and the patient-facing view.

→ Open ClearPath prototype

Opens in a new tab · No login required

Business case

Three ROI levers: throughput, outcomes, workforce retention

2–4 hrs

Avg avoidable discharge delay per patient

~$900

Excess bed-day cost per delayed discharge

Max CMS readmission penalty (HRRP)

20–25%

Annual coordinator turnover rate

Throughput ROI

A 400-bed hospital at 85% occupancy discharges ~70 patients/day. Recovering 1 hour of delay across 20% of discharges = 14 hours of bed capacity recovered daily. At near-capacity, recovered bed-hours translate directly to additional admissions.

Clinical outcomes ROI

Poor discharge coordination is a primary driver of 30-day readmissions, which trigger CMS HRRP penalties. Improved discharge adherence and plain-language patient updates directly reduces readmission risk and penalty exposure.

Workforce ROI

Reducing status-chasing from 60–70% to ~30% of a coordinator's shift — roughly 3–4 hours/day — reduces cognitive burden. Measured via monthly coordinator NPS and 6-month retention tracking.

Deployment plan

Weeks, not months — gated on outcomes

Shadow mode

Weeks 1–4

Read-only access — no EHR writes, no workflow changes

AI runs in background on one med-surg unit

Measure blocker ID accuracy vs. coordinator ground truth

Gate: 80%+ accuracy on primary blockers

Recommendations

Weeks 5–10

Surface recommendations to coordinators (read-only)

Measure override rate by blocker category

Activate alerting for 4+ hour escalations

Gate: coordinator adoption above 60%

Integration + scale

Weeks 11–20

FHIR API integration for real-time structured data

Expand to 2–3 units with prompt refinement

Phase 3+: patient-facing companion layer

Measure: LOS delta, readmission, retention

Alert fatigue

Too many low-confidence outputs erode trust silently. Mitigation: conservative thresholds in pilot — surface fewer, higher-confidence recommendations only.

Liability clarity

When AI informs a decision that goes wrong, accountability must be unambiguous: the coordinator, always. Explicit in training, UI design, and clinical governance.

Equity

A model trained on academic medical center notes may underperform in safety-net hospitals. Pilot must include a safety-net institution from day one — not after scaling.

I build products by starting with the people who have to use them