Kurta Squad

We use Muse Spark (muse-spark-1.3) through the Meta Model API in two places. First, it drafts the clinical report for each MRI study: it receives the image plus the classifier output and returns structured JSON (strict JSON schema), where every claim lists the evidence keys that support it. Second, it answers questions about the study in a streamed chat. Muse Spark never classifies; a separate hybrid classifier does. A citation guard then checks each claim against the classifier's numbers and marks it Verified or Not verified. We set a reasoning effort per use (low for the report, minimal for the chat), bound tokens per request, and added a daily token budget and a requests-per-minute cap in the backend. The model sits behind an LLM port, so an open Llama model can replace it.