I am having an aws ecs service having 2 tasks running on t2.xlarge. The service is behind application load balancer.

Service contains an http api deployed using gunicorn + uvicorn + fastapi (4 workers) which takes image of size 112 * 112 * 3 as input and uses tf lite model to predict. This processing takes less than 200 ms while the complete turn around time of API is more than 12 seconds. Can anyone suggest how this time can be improved?