🏗️ Yorkie Bakery – System Architecture Overview

This page provides a technical high-level overview (HLD) of the Yorkie Bakery platform, including backend architecture, AI/ML pipeline, database design, and deployment infrastructure. This is intended for recruiters and engineers evaluating my work.

1. High-Level Architecture Diagram

                ┌────────────────────────┐
                │      React + Vite      │
                │  (AI Demo Frontend)    │
                └───────────┬────────────┘
                            │  REST + JSON
                            ▼
                  ┌──────────────────────┐
                  │       FastAPI        │
                  │  Auth / Menu / AI    │
                  └──────────┬───────────┘
                             │
     ┌─────────────┬─────────┼─────────┬──────────────┐
     ▼             ▼         ▼          ▼               ▼
PostgreSQL     ChromaDB   OpenAI    AWS S3 (Music)   AWS SES
(Menu/Users)   Vectors    Vision /   Audio Store    Email Verify
                         Embedding
  

2. Backend Architecture (FastAPI)

The backend is a modular FastAPI application deployed on AWS EC2 using Docker Compose. It includes:

• 🔹 **User authentication** (register, login, session cookies, email verification)
• 🔹 **Menu management** (admin CRUD, S3 image uploads)
• 🔹 **Music streaming** (S3 signed URLs)
• 🔹 **AI endpoints** for RAG, filters, and image matching

API Modules

• auth/ – login, register, session, email verification
• menu/ – product CRUD, pagination, filters
• ai/demo/ – AI pipeline (LLM filters + vector search)
• ai/vision/ – image upload → OpenAI Vision model

3. AI Menu Recommendation Pipeline

The AI system uses a two-stage hybrid retrieval approach combining LLM semantic understanding with vector similarity search and strict backend filtering.

Pipeline Steps

1. **User sends text or image** from the React demo.
2. FastAPI calls **OpenAI** to extract:
   • category (pastry / dessert / drink)
   • origin (japanese / french / thai / …)
   • flavor profile (sweet, fruity, chewy, creamy)
   • price range
3. Use **ChromaDB vector store** to retrieve the top 50 candidate menu items.
4. Apply **server-side filtering** with `apply_all_filters()`.
5. Rank + return the final result set to React.

Vision Image Flow

Upload Image → FastAPI → OpenAI Vision → Tags/Flavor → Vector Search → Recommendations
  

4. Database Schema

PostgreSQL Tables

• user_account – auth, email verification
• menu_item – title, description, price, tags
• music_track – title, s3_url
• orders + order_items (future extension)

ChromaDB Vectors

The embedding store uses OpenAI embedding vectors for each menu item:

{ id, title, tags, flavor_profiles, embedding_vector[] }
  

5. Deployment Architecture

• 🐳 **Docker Compose** (FastAPI + Postgres + ChromaDB)
• 🐙 Hosted on **AWS EC2**
• 🌐 Domain: yorkiebakery.com
• 🧪 Staging: beta.yorkiebakery.com
• 📧 AWS SES – email verification & password reset
• 🪣 AWS S3 – music tracks + images

6. Future Enhancements

• ⭐ User preference profiles (learn from history)
• ⭐ Collaborative filtering recommendations
• ⭐ Better ranking using LLM scoring + vector similarity
• ⭐ Admin analytics dashboard