Abstract: The present disclosure provides a computer-implemented method for reducing latency by transitioning from a Large Language Model (LLM) to a specialized Machine Learning model for performing operations within an application. The method includes receiving a user instruction specifying an operation, executing the operation based on LLM output, determining a performance metric for the LLM, identifying a specialized Machine Learning model, training the specialized model using stored data, deploying the specialized model within a software development kit, and redirecting subsequent instructions to the specialized model. The training process involves building a labeled dataset mapping user instruction to API calls or parameter sets and initializing the specialized model with pre-trained or random parameters. The method enables efficient handling of application-specific tasks by leveraging the initial capabilities of the LLM and subsequently transitioning to more focused, specialized models for improved performance.

Abstract: A system and a method for automatically discovering and managing actions in an application is disclosed. The system includes a data ingestion layer for receiving application data from multiple sources, a scanning and systematic traversal engine for interacting with UI elements and capturing network calls, an action mapping and generation module for correlating UI actions with API calls and categorizing actions, an AI-driven icon and description generator for creating visual representations and textual descriptions of actions, a user interface for displaying and modifying discovered actions, and a continuous monitoring component for triggering re-scanning based on coverage metrics, error detection, or version updates. The system employs synthetic data generation and AI-driven exploration to uncover hidden or undocumented APIs, enabling comprehensive mapping of an application's capabilities at the API level.