Talk to an AI medical assistant by voice or text, book verified doctors, and manage your entire health record — in one calm, secure place.
Live walkthrough: sign-in · patient dashboard · AI consultation (Nepali) · appointments · health records · doctor portal
MeroDaktar ( “My Doctor” ) is a full-stack telemedicine platform that brings primary-care triage to anyone with a phone. A patient describes how they feel — by voice or by typing — and a medical-grade AI assistant runs a structured symptom interview, flags urgency, remembers the conversation, and can hand off to a real, verified doctor for an appointment. Every interaction is captured in a proper Electronic Health Record (EHR).
It is not a chatbot demo. It is a layered, tested, production-shaped system: 65 REST endpoints, a clean routes → services → repositories → models backend, a pluggable AI engine, Redis-backed conversation memory with semantic search, and a complete doctor portal.
| 🎙️ Voice-first triage | 🧠 Medical-grade AI | 🔁 It remembers |
| Speak your symptoms in Nepali or English — Gemini transcribes, the AI replies in your language, and reads answers aloud. | Pluggable engine via one env var — Gemini, MedGemma, or OpenAI — same prompts, swappable backend. | Redis session memory + semantic vector search mean no repeated questions across a consultation. |
| 🩺 Real doctor handoff | 📋 Full EHR | 🛡️ Built to last |
| Live availability, time-slot booking, and a complete doctor dashboard with encounter notes. | Vitals, allergies, medications and a chronological encounter timeline per patient. | 49 automated tests · 91% coverage, JWT auth, Dockerised infra. |
| Metric | Value |
|---|---|
| REST API endpoints | 65 (27 POST · 24 GET · 8 PUT · 5 DELETE · 1 PATCH) |
| Backend Python | 7,630 lines across 68 modules |
| Frontend TypeScript / React | 6,536 lines — 9 screens + a shared design system |
| Feature modules | 11 (auth, chat, appointments, EHR, schedules, reports, doctors, dashboard, voice ASR, users, admin) |
| Data models | 8 (User, Doctor, Appointment, Schedule, Encounter, Consultation, EHR, Report) |
| Automated tests | 49 — 31 unit · 15 integration · 3 end-to-end |
| Test coverage | 91% |
| AI engines | 3 — MedGemma · Gemini 2.0 Flash (ASR) · OpenAI (gpt-4o-mini + embeddings) |
| Contributors | 3 |
| Total codebase | ~15,500 lines |
 🏠 Patient dashboard — health hub with quick actions, recent consultations & upcoming visits |
 💬 AI consultation — voice + text symptom triage with conversation history |
 📅 Appointments — browse doctors, pick a slot, book & manage |
 📋 Health records (EHR) — vitals, allergies, medications & encounter timeline |
 🩺 Doctor portal — overview, appointments, patients & schedule management |
 🔐 Separate doctor portal — dedicated, JWT-secured sign-in |
The flows below are live in the app today.
Sign up / Log in ─▶ Dashboard ─▶ "AI Consultation"
│ │
│ ┌─────────────────┴─────────────────┐
│ 🎙️ Speak symptoms ⌨️ Type symptoms
│ └─────────────────┬─────────────────┘
│ ▼
│ AI triages → asks follow-ups → flags urgency 🟢🟡🔴
│ ▼
│ "This looks moderate — let's book a doctor."
▼ ▼
Health Records (EHR) ◀── Book Appointment ─▶ Doctor reviews & adds encounter notes
1. Voice or text consultation. The patient opens AI Consultation, taps the mic (or types), and describes their symptoms. Gemini transcribes speech in real time; the AI assistant conducts a guided symptom interview and classifies urgency as routine / moderate / emergency.
2. Memory that follows the conversation. Redis stores the session and a semantic index of prior turns, so the assistant never asks the same question twice and keeps full patient context (demographics, history, EHR) in every prompt.
3. Seamless doctor handoff. When the patient needs a clinician, they book a real appointment against a doctor's live availability. Doctors get a full dashboard — appointments, patient records, and per-visit encounter notes that flow back into the EHR.
4. One health record. Vitals, allergies, medications and a chronological encounter timeline live in the patient's EHR, viewable any time.
flowchart TB
subgraph Client["🖥️ Frontend — React 18 · TypeScript · Vite · Tailwind"]
UI["Patient & Doctor SPA<br/>voice recorder · chat · EHR · scheduling"]
end
subgraph API["⚙️ FastAPI — layered backend (65 endpoints)"]
direction TB
R["Routes /api/v1<br/>auth · chat · appointments · ehr · schedules · reports · doctors · admin"]
S["Services<br/>business logic · AI orchestration"]
Repo["Repositories<br/>data access"]
M["SQLAlchemy Models"]
R --> S --> Repo --> M
end
subgraph AI["🧠 AI Engine (pluggable via AI_BACKEND)"]
MG["MedGemma<br/>medical LLM"]
OAI["OpenAI<br/>gpt-4o-mini + embeddings"]
ASR["Gemini 2.0 Flash<br/>speech → text"]
end
subgraph Data["💾 Data layer"]
PG[("PostgreSQL<br/>system of record")]
RD[("Redis<br/>session memory + vector search")]
end
UI -->|JWT REST| R
UI -->|🎙️ audio| ASR
S --> MG
S --> OAI
S --> RD
M --> PG
ASR --> S
sequenceDiagram
actor P as Patient
participant FE as Frontend (SPA)
participant API as FastAPI
participant G as Gemini ASR
participant LLM as MedGemma / OpenAI
participant R as Redis
P->>FE: 🎙️ Records symptoms (MediaRecorder)
FE->>API: POST /api/v1/speech/transcribe (audio)
API->>G: transcribe
G-->>API: transcript
API-->>FE: text
FE->>API: POST /api/v1/chat/session/{id}/message
API->>R: load session memory + semantic context
API->>LLM: prompt (symptoms + full patient context)
LLM-->>API: triage + follow-up + urgency
API->>R: persist turn + embeddings
API-->>FE: response
FE-->>P: 🔊 reads answer aloud (TTS) + urgency badge
| Area | Patient | Doctor |
|---|---|---|
| AI consultation | Voice + text symptom triage, urgency flags, conversation history | — |
| Memory & context | Redis session memory, semantic search, full EHR context in prompts | — |
| Appointments | Browse doctors, live slots, book / cancel | Manage availability, confirm / complete, cancel |
| EHR | Vitals, allergies, medications, encounter timeline | Read patient records, write encounter notes |
| Reports | AI-generated symptom-assessment reports | Patient-linked reports |
| Auth & security | JWT login / registration | Separate doctor portal & JWT |
| Dashboards | Health summary, recent consultations, upcoming visits | Stats, schedule, patient & appointment management |
Frontend · React 18 · TypeScript · Vite · Tailwind CSS · React Router · Axios · Heroicons
Backend · FastAPI · Uvicorn · SQLAlchemy · Pydantic v2 · python-jose (JWT) · passlib + bcrypt
Data · PostgreSQL · Redis (caching, session memory, vector search)
AI · Google MedGemma · Gemini 2.0 Flash (ASR) · OpenAI gpt-4o-mini & text-embedding-3-small · google-genai
Tooling · Docker Compose (postgres · pgadmin · redis) · pytest · pytest-asyncio · pytest-cov
The frontend was rebuilt on a bespoke “Health-Tech Gradient” design system — deep-navy surfaces, violet→cyan gradients, glassmorphism and soft glow, with Space Grotesk display type over Inter. Everything renders from shared primitives so the whole product feels cohesive:
src/lib/ui.tsx—Button,Card,Input,Select,Field,Badge,StatCard,Spinner,Avatar,PageHeader,EmptyStatesrc/components/layout/AppLayout.tsx— responsive sidebar + top-bar shell for patient & doctor portalssrc/components/layout/AuthLayout.tsx— split hero + form for sign-in / sign-uptailwind.config.js+src/index.css— design tokens, animations and glass utilities
merodaktar/
├── app/ # FastAPI backend (7,630 LOC · 68 modules)
│ ├── api/v1/ # 11 route modules · 65 endpoints
│ ├── services/ # business logic + AI orchestration
│ ├── repositories/ # data-access layer
│ ├── models/ # 8 SQLAlchemy models
│ ├── schemas/ # Pydantic request/response models
│ ├── core/ # security, middleware, exceptions
│ ├── config/ # settings + database
│ └── migrations/ # schema migrations
├── frontend/ # React + TypeScript SPA (6,536 LOC)
│ └── src/
│ ├── components/ # 9 feature screens
│ ├── components/layout/ # AppLayout (sidebar shell) + AuthLayout
│ └── lib/ui.tsx # shared design-system primitives
├── tests/ # 49 tests — unit · integration · e2e
├── docker-compose.yaml # postgres + pgadmin + redis
└── requirements.txt
docker compose up -d # starts postgres, pgadmin, redispython -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp app/.env.example app/.env # add your DB url + AI keys (MEDGEMMA / OPENAI / GEMINI)
uvicorn app.main:app --reload # http://localhost:8000 · docs at /docsSwitch AI engines with a single env var:
AI_BACKEND=gemini(chat + embeddings + voice on one key),AI_BACKEND=openai, orAI_BACKEND=medgemma.
cd frontend
npm install
npm run dev # http://localhost:5176pytest # 49 tests
pytest --cov=app # with coverage (~91%)MeroDaktar is the work of three engineers. Every contributor's original commits are preserved in this branch's history.
| Contributor | Focus | |
|---|---|---|
| 🧱 | Utshav Paudel | Project lead & architecture — core platform, layered backend refactor, EHR, appointments, doctor portal, MedGemma integration, test suite, full UI redesign |
| 🎙️ | Sumit Thokar | Voice AI — Gemini speech transcription (ASR) and in-chat voice-input recording |
| 🧠 | Shishir Bhattarai | Chat memory & medical-chat API, appointment management and dashboard UX |
# every contributor's work is in the history of this branch:
git shortlog -sneReleased for educational and portfolio purposes. © 2025 the MeroDaktar team.
MeroDaktar — your health, one tap away.