Middleware builds API layer that fine-tunes LoRA adapters from user corrections to keep task-specific models current
The Middleware team announced a waitlist for an API layer that improves models on repetitive tasks like invoice extraction and support replies by learning from user corrections. LoRA adapters refresh every few minutes, with updates kept inexpensive and support spanning multiple model families. Integration requires only changing the API base URL, and users can keep existing providers alongside task-trained models.