Doesn't that look like an interesting approach for highly optimized embeddings?