Overview

The Daycare AI Platform is an AI-native childcare operations system designed to replace manual administrative workflows (like tapping through apps or hiring virtual assistants) with a voice-first, high-accuracy pipeline. The platform allows daycare teachers to log daily events using voice memos and photos via WhatsApp. This unstructured data is processed, structured, validated, and pushed through a multi-tier review system before generating cohesive, AI-driven daily narrative reports for parents.

The Problem It Solves

Independent daycare directors ($300K–$750K ARR) rely on consumer-grade tools (like Brightwheel) which require teachers to manually log events on tablets (disrupting the classroom). Alternatively, they hire offshore VAs to review photos and create logs, which is expensive, slow, and prone to context loss.

This platform acts as an automated, highly-accurate operations layer. It eliminates the manual data entry bottleneck, guarantees structured data via LLM extraction, and uses a human-in-the-loop (HITL) model to prevent AI hallucinations from ever reaching parents.

System Architecture & Pipeline

The system is a standalone platform with full stack ownership (no dependency on legacy tools).

1. Ingestion Layer (WhatsApp Business API via Twilio)

   • Teachers interact exclusively via a WhatsApp Bot. They send voice memos (.ogg/.mp4 up to 90s) and photo batches.
   • Natural language context (e.g., "Jason just ate all his mac and cheese") serves as the raw input.

2. AI Extraction & Structuring Layer (FastAPI Backend)

   • Speech-to-Text (STT): OpenAI Whisper (V1) is used for rapid transcription. (Future iterations plan to use AssemblyAI for better proper noun resolution).
   • Entity Extraction (LLM): GPT-4o processes the transcript at temperature=0 to ensure deterministic outputs. It extracts data strictly matching predefined Pydantic schemas (JSON structure).
   • Strict Validation: The extracted JSON maps to strictly-typed entities (e.g., MEAL, NAP, DIAPER, BILLING_LATE_PICKUP). Parsing failures or ambiguous events default to needs_review: true.

3. Three-Tier Human-in-the-Loop (HITL) Review System

   • Teacher Review: High-confidence parsed events are sent back for a single-tap approval.
   • Director Review: Low-confidence, incident-related, or billing-related events are automatically flagged for the Director Console.
   • Parent View: Parents only see data after explicit human approval.

4. Presentation & Narrative Generation

   • At the end of the day, an automated pipeline groups all approved events per child.
   • GPT-4o synthesizes these discrete events into a warm, personalized 120-250 word narrative paragraph.
   • Parents access these reports via a Next.js-based web portal using magic links (no app install required).

5. Billing & Monetization Module

   • Voice memos containing billing intents (e.g., "Marcus was picked up 45 mins late") trigger a specific billing schema extraction.
   • Admins approve the extraction, generating a PDF invoice processed via Stripe Connect.

Tech Stack

• Backend: Python 3.11, FastAPI, Uvicorn, Pydantic (Strict Data Validation).
• Database: PostgreSQL (Multi-tenant scoped by center_id), SQLAlchemy + Alembic (Migrations).
• Frontend (Admin/Teacher): React 19 (PWA, Vite, TailwindCSS) for real-time review queues.
• Frontend (Parent Portal): Next.js for SSR, magic-link auth, and SEO/performance.
• AI/ML: OpenAI API (Whisper + GPT-4o).
• Integrations: Twilio (WhatsApp API), Stripe Connect (Payments).

Core Engineering Guardrails

• Zero Hallucination Tolerance: Strict schema adherence using temperature=0. Any deviation fails validation and is sent to the human review queue.
• Data Isolation: All transactional queries are scoped by center_id at the DB level ensuring multi-tenant data privacy.
• Compliance: Built with COPPA compliance in mind (media EXIF stripping, secure buckets, strict authorization models).

Local Development Setup

1. Backend Environment:

   python -m venv venv
   source venv/bin/activate
   pip install -r backend/requirements.txt

2. Database Migrations:

   alembic upgrade head

3. Run FastAPI Server:

   cd backend
   uvicorn main:app --reload

4. Run Frontend Clients:

   # Admin & Teacher Console
   cd frontend/console
   npm install && npm run dev
   
   # Parent Portal
   cd frontend/parent
   npm install && npm run dev

Documentation

Full product requirements and structural overviews are detailed in docs/PRD.md and docs/legal_PRD.md.