1 / 13
Space · Arrows · Swipe
HB
HackPrinceton 2025 · Demo
HealthBridge
WhatsApp Health Navigation

Grounded · Multilingual · Action-Oriented

TypeScript + Node.js WhatsApp · Baileys Gemma · K2 Think V2 RxNorm · openFDA Docker · HF Spaces

HealthBridge supports health decisions — it does not diagnose or replace a licensed provider.

Cue: "An engineered health navigation system — not a chatbot wrapper — where millions already are: WhatsApp."

FEAR
The Problem

US Healthcare
Is a Maze of Fear

Millions delay care — not from lack of access — but from fear, language barriers, and confusion.

  • Afraid to call 911
    Cost anxiety or immigration fear — people wait dangerously long.
  • Unreadable paperwork
    Discharge papers, EOBs, and prior auth letters are inaccessible.
  • Medication confusion
    Taking prescriptions without knowing dangerous interactions.
  • No provider direction
    Sliding-fee clinics exist nearby — nobody knows they're there.
41%
of US adults delayed or skipped care due to cost last year
67M+
people in the US speak a language other than English at home
80%
of medical errors linked to miscommunication

Cue: "The problem isn't coverage — it's navigation. People freeze when they don't understand the system."

WHO
Audience

Built for Anyone
Navigating US Healthcare

Not just immigrants — anyone uncertain or overwhelmed. WhatsApp already has 2B+ users. We meet them there.

👨‍👩‍👧

Middle-Income Families

Too much for Medicaid, too little for surprise bills. Needs clear cost guidance.

🌍

Immigrants & Newcomers

Multilingual triage, FQHC navigation, safe-access guidance regardless of status.

🎓

Students & Young Adults

Newly off parents' insurance. Alone in medication and urgent care decisions.

⚙️

Gig & Hourly Workers

No employer plan. Need fast triage before deciding if the ER is worth the bill.

🧑‍⚕️

Caregivers

Managing medications and follow-ups for elderly parents with complex needs.

🏙️

Underserved Communities

Low health literacy, low system trust. Need a non-judgmental guide in their language.

Cue: "140M Americans have no family doctor they can call at 11pm. We built HealthBridge for all of them."

FEAT
Product Overview

Six Features.
One WhatsApp Message.

Every feature ends with a concrete action — not a wall of caveats.

🚨

Emergency Override

Deterministic regex in 10+ languages. 911 directive — no LLM latency in the critical path.

💊

Medication Intelligence

Photo a pill bottle. Gemma extracts, FDA checks interactions, K2 explains clearly.

🏥

Symptom Triage

6-step flow via MedlinePlus + K2 reasoning. ER, urgent care, or wait at home?

📍

Provider Finder

ZIP-based HRSA/FQHC lookup. Nearest sliding-fee clinic in under 2 seconds.

Medication Reminders

Schedule reminders, reply "taken" to confirm. For caregivers and chronic patients.

🛡️

Safe-Access Rights

Patient rights and protections for those afraid to seek care due to cost or status.

Cue: "Every feature is an action, not information. We guide to the next step — we don't just answer."

FLOW
End-to-End Demo

A Real User Journey
in Three Messages

01
Text Query
"Dad takes metformin + lisinopril. New ibuprofen — okay?" → router classifies as medication intent.
02
Image Sent
User photographs prescription. Gemma extracts drug, dose, frequency → matched to RxCUI via RxNorm.
03
FDA Check
openFDA FAERS: NSAIDs + ACE inhibitors = renal risk. 234 adverse reports. Safety score: 🟡 YELLOW.
04
K2 Reasoning
Chain-of-thought over grounded data + patient context → plain-language guidance + one clear action.
05
Clinic Found
User sends ZIP. HRSA lookup returns nearest sliding-fee clinic with hours + phone in <2 seconds.

Cue: "From confusion to clinic referral in one conversation. No apps, no logins, no dead ends."

ARCH
System Architecture

Five Layers of
Engineered Intelligence

DETERMINISTIC
Layer 1
Safety + Routing
Emergency keyword detection (10+ langs). Intent classification. No LLM in critical path.
GEMMA 3 27B
Layer 2
Multimodal Extraction
Image/PDF → structured JSON. Drug, dose, prescriber, doc type via OpenRouter.
GROUNDED
Layer 3
Data Grounding
RxNorm → RxCUI. openFDA labels. FAERS adverse signals. MedlinePlus. HRSA FQHC.
K2 THINK V2
Layer 4
Deep Reasoning
Chain-of-thought over grounded data. Follow-up context. Plain-language guidance.
ACTION
Layer 5
Response + Action
WA-formatted reply. Safety label. Clinic referral. Reminder scheduling.
Fastify · TypeScript Baileys WA transport HMAC session security Ephemeral memory Docker · HF Spaces Zero PHI persistence

Cue: "Emergency? Deterministic. Document? Gemma. Safety? FDA. Guidance? K2. Each layer has one job."

AI
Layer 2 · Extraction

Gemma Multimodal
Extraction Layer

google/gemma-3-27b-it via OpenRouter

Users send real health documents. Gemma converts visual data into structured JSON that downstream layers reason on — no manual re-typing.

  • Rx
    Prescription Labels

    Drug name, dose, route, frequency, prescriber. Normalized to RxCUI downstream.

  • 💊
    Pill Bottle Photos

    OCR handles glare, rotation, partial occlusion via vision model.

  • 📋
    Discharge Papers

    Diagnosis codes, follow-up instructions, prescribed medications from hospital paperwork.

  • 🔬
    Lab Results & EOBs

    Numeric values + reference ranges; coverage decisions + appeal deadlines surfaced clearly.

Output JSON Schema

{ drugName: string,
  dose: string,
  frequency: string,
  prescriber: string,
  docType: enum,
  confidence: 0–1 }

PDF Text Path

Direct text extraction — no vision cost. Regex-assisted parsing. Falls back to Gemma vision if text layer is absent.

Cue: "A user doesn't type drug names — they photograph the bottle. Gemma reads it. We normalize it. K2 reasons on it."

FDA
Layer 3 · Grounding

FDA + RxNorm
Medication Intelligence

RxNorm Drug Normalization

1
Drug name → rxnorm/rxcui lookup. Handles brand names, generics, misspellings.
2
Returns canonical RxCUI — stable key for all downstream FDA queries.
3
openFDA label lookup → indications, contraindications, boxed warnings, pregnancy category.
4
Each drug pair checked against FAERS adverse-event co-occurrence database.
5
Death-report signal count extracted. High signal → elevated safety tier sent to K2.

⚠ FAERS Adverse Event Signals

A
Query: /drug/event.json?search=patient.drug.medicinalproduct
B
Returns total adverse reports, serious outcomes, death count for the drug pair.
C
Signal thresholds map to safety tier → passed as context to K2.
🟢 GREEN
No signals
🟡 YELLOW
Moderate risk
🔴 RED
Deaths reported

Cue: "Real FDA data, normalized, scored, grounded — before we touch the reasoning layer."

K2
Layer 4 · Reasoning

K2 Think V2
Deep Reasoning Layer

K2 receives a structured context packet — grounded FDA data, extracted document fields, conversation memory — and produces one clear next step.

🧠
Chain-of-Thought over Grounded Data
Receives safety score, drug label warnings, FAERS counts. Reasons step-by-step before output.
💬
Conversation Memory
Ephemeral context window maintains prior exchanges within a session.
🎯
Action-Oriented Outputs
Every response ends with a concrete action: "Ask your pharmacist", "Go to urgent care", "Call this clinic".
🌐
Multilingual Output
Responds in the user's detected language. Spanish, Mandarin, Hindi, Portuguese supported.

Reasoning Chain Example

01 — Context Loaded

Metformin + lisinopril + ibuprofen. Safety score: YELLOW. FAERS signal: 234 reports.

02 — Mechanism

NSAIDs reduce renal prostaglandins → impair ACE inhibitor efficacy → potential acute kidney injury.

03 — Output

🟡 Plain explanation. "Call your pharmacist — ask about an NSAID alternative."

Cue: "K2 doesn't just answer — it reasons. Grounded in real FDA data, it thinks before it speaks."

SAFE
Trust & Safety

Safety, Privacy
& Guardrails

🔐 Session Security

  • Phone numbers HMAC-hashed. Never stored in plaintext.
  • Ephemeral in-memory sessions — no persistent PHI vault.
  • HTTPS enforced. Input sanitization on all inbound messages.

🚦 Rate Limiting & Abuse

  • Per-session rate limiting on all expensive AI calls.
  • Input length caps and content sanitization before any LLM pass.
  • Emergency override bypasses rate limits — 911 always gets through.

🚨 Emergency Override

  • Deterministic regex — zero LLM latency in the crisis path.
  • Triggers: chest pain, stroke, suicidal ideation, severe bleeding + more.
  • 10+ languages. 911 directive + crisis hotline returned in <50ms.

⚖️ Clinical Guardrails

  • System prompt enforces "support, not diagnose" on every LLM call.
  • All medication guidance includes "confirm with pharmacist" directive.
  • Triage references MedlinePlus — a trusted government source.

Cue: "Emergency detection is deterministic — it doesn't wait for an LLM."

TECH
Infrastructure

Integrations, Deployment
& Operations

Transport

Baileys + WhatsApp

Open-source WA multi-device library. QR pairing on /qr. Inbound router + media download.

AI Models

OpenRouter Gateway

Unified inference for Gemma 3 27B (vision) and K2 Think V2 (reasoning). Provider-agnostic.

Data Sources

Federal Health APIs

RxNorm, openFDA, FAERS, MedlinePlus, HRSA FQHC locator. All public — zero licensing cost.

Runtime

Node.js + Fastify

TypeScript throughout. Fastify high-throughput routing. Zod schema validation. Structured logging.

Deployment

Docker + HF Spaces

Single-image Docker build. Hugging Face Spaces hosting. One-command deploy.

Key Routes

API Endpoints

  • GET /health — WA connection status
  • GET /qr — device pairing QR
  • POST /send — outbound dispatch

Cue: "Every dependency is public or open-source. Zero licensing cost. One docker run and it's live on HF Spaces."

WHY
Why HealthBridge Wins

Impact &
Differentiation

HealthBridge

Grounded in real FDA data — drug safety never hallucinated
Emergency override is deterministic — crisis speed, multilingual
Multimodal: reads your actual prescription, not what you type
Zero app download — meets users on WhatsApp they already use
ZIP → nearest clinic in <2s — not a vague Google search

Generic Health AI

LLM-only: drug info hallucinated from training data
Emergency detection via prompt — latency risk in crisis
User must read and retype all information manually
App download, account creation, onboarding friction
Provider search is a link dump, not ZIP-personalized
2B+
WhatsApp users
1,400+
FQHC sites via HRSA
$0
Licensing cost
10+
Emergency languages
5
Intelligence layers

Cue: "This isn't ChatGPT on WhatsApp. It's a layered clinical navigation system with real grounding and triage logic."

HB

HealthBridge

Grounded · Multilingual · Deterministic where it matters

Try the Live Demo View GitHub Repo

HealthBridge is a decision-support tool. It does not diagnose, prescribe, or replace licensed medical providers.

Cue: "Message +1 (920) 489-5575 on WhatsApp. Ask about a medication. See why this is engineered navigation — not a chatbot."