Hey everyone!
This is my first post and I wanted to start sharing this project with you all.
I've been experimenting with audio embeddings recently to see if I could build a self-hosted search tool for music.
The result is a prototype called (for now) Agnostic Intelligence Layer. It's a semantic audio search engine designed to run entirely offline without reliance on external cloud APIs.
The Stack & Architecture
I wanted something fast and efficient, so I decided to mix Java and Python:
- Java Quarkus: Handles the core engine pipeline and container efficiency.
- Python: Manages the actual AI heavy lifting using CLAP neural networks to extract audio features into 512-dimensional vectors.
- PostgreSQL + pgvector: Stores the vectors and finds acoustically similar tracks using cosine similarity.
- MinIO: Handles fast and easy local storage.
Everything starts up via a single docker-compose, so you don't have to waste time configuring external services.
Why I'm sharing this
It's still a prototype, but the core pipeline works fine, and I'm going to continue working on it. I wanted to share it early to get some eyes on the code and see if the architecture makes sense to other devs.
The repository has a quick start guide if you want to check out the code or test it locally:
👉 https://github.com/BothBasilisk/agnostic-audio-engine.git
Feel free to leave your thoughts, critiques, or any tips on how to improve the pipeline!