AI architecture here is how assistant, copilot, knowledge, and admin share models, data, and boundaries without becoming one blob.
In my daily workflow
- I draw boundaries: public read, auth write, agent tools per surface.
- Shared AI modules live in `features/ai`; product logic stays in feature folders.
- I avoid circular deps between retrieval and generation layers.
- I revisit the diagram when adding a new agent or data source.
How it makes me work smarter
Clear architecture stops duplicate routers and leaked keys. One provider module, one embedding pipeline, many consumers. New surfaces (gallery copilot, job-fit) plug in instead of forking stacks.
My setup
- Feature folders: ai-assistant, copilot, knowledge-base, admin
- Shared provider router and model config
- Supabase as system of record for content and vectors
- LangGraph for copilot; lighter chains for assistant
On this portfolio
Public pages, admin CMS, assistant, and copilot share Supabase content but use different auth and tool scopes. The AI-first gallery documents the patterns; the repo implements them as separated feature modules.
