The Architecture of My AI Lab

This is the technical backbone of my AI Lab: how I set up RAG search, voice assistant, and prompt injection testing in a Next.js environment with Edge Functions and Supabase.

This post walks through how I built architecture, and where it fits in the rest of my work.

This architecture builds on the same patterns I developed for museum kiosks and government platforms, where stability and observability were just as important as features.

My AI Lab contains three major projects:

1. RAG Knowledge Search
2. Voice Assistant
3. Prompt Injection and Safety Lab

They all share a unified architecture.

The foundation

The AI Lab uses:

• Next.js API routes
• Supabase for persistence
• OpenAI for model calls
• pgvector for vector search
• Shared utilities and error handling

Shared design principles

• Streaming responses
• Stateful UI where needed
• Server side input validation
• Structured metadata returned with every model call

RAG architecture

The RAG system uses:

• Document ingestion panel
• Chunking with overlap
• Embeddings stored in pgvector
• SQL similarity search
• Streaming final answers
• Citation panels

Voice assistant architecture

The voice assistant uses:

• Browser SpeechRecognition
• Whisper fallback endpoint
• OpenAI TTS for output
• Tool calling for weather, timezone, search, and math
• Session memory

Prompt injection lab architecture

The security lab uses:

• Level configuration objects
• Backend guards per level
• Output validators
• Custom UI per level
• LocalStorage progress

Why this matters

This system combines design multiple interoperable AI products inside one codebase in a production environment.

Keep exploring

From here you can:

• See how I applied similar patterns on the American Battlefield Trust map projects