🫧 Murmur

Murmur is a real-time support tool for caregivers of neurodivergent children. It guides users through emotionally intense moments by helping them identify observable behavior and receive trauma-informed, empathetic responses generated via AI.

---

🧠 Purpose

Murmur aims to improve caregiver-child communication by:

• Guiding caregivers through structured, observation-based prompts
• Generating helpful, real-time language and behavioral insights using LLMs
• Supporting co-regulation and empathy during difficult moments

---

⚙️ How It Works (MVP)

1. Onboarding: Users answer a short questionnaire about their child to create a personalized BaseContext (e.g., age, diagnosis, communication style, tone preference).
2. Observation Selection: During a challenging moment, the user selects from a curated list of observable behaviors (e.g., "They are covering their ears"). These come from a static, Firestore-hosted Situational Context Tree (SCT).
3. Prompt Generation: The app combines the BaseContext and selected observations into a structured prompt, which is sent to OpenAI's API.
4. AI Response: The backend returns a brief suggested caregiver response and a follow-up question, which are displayed in the app.
5. (Optional): The user can provide feedback or request a refined suggestion.

---

🧱 Tech Stack

🖥️ Frontend (console/)

• React with Vite
• TypeScript
• Tailwind CSS
• Axios (for API requests)
• Loads and displays:
  • Observation prompts (from Firestore)
  • User responses (from backend API)
  • Calming and accessible UI

🔧 Backend (server/)

• Python + FastAPI
• Receives base context + user-selected observations
• Generates and sends a structured prompt to the OpenAI API
• Returns: suggested caregiver response + follow-up question

📦 Database (Firebase Firestore)

• BaseContext: stored per user in /users/{userId}/baseContext
• Situational Context Trees (SCTs): static, app-wide trees stored in /contextTrees/, loaded by the frontend
• No per-session or chat data is stored for MVP

---

🛠 Developer Notes

• The situational context tree is read-only and should be structured as a nested JSON object with:
  • prompt: the current question
  • options: an array of choices, each with a label, value, and optional followUp
• The frontend builds a list of selected observations which, together with the base context, forms the basis for the AI prompt
• All LLM interactions are handled by the backend to avoid exposing the API key

---

💻 To Run Client

cd client
npm install
npm run dev

🛜 To Run Server

cd server
uvicorn main:app --reload

---

🗂️ Data & State Management

🔹 Base Context (User-specific)

During onboarding, each user submits a baseContext that is stored in Firestore at:

This includes: /users/{userId}/baseContext

{
  "caregiverName": "Alex",
  "childName": "Sam",
  "childAgeRange": "4–6",
  "diagnoses": ["autism"],
  "communicationStyle": "non-verbal",
  "language": "English (US)"
}
##youtube link: https://youtu.be/yGCSw11w2pc