AAI_2025_Capstone_Chronicles_Combined

‭ResolveAI‬

‭be effectively leveraged through proper encoding and feature engineering to enhance model‬ ‭performance without causing significant multicollinearity issues.‬

‭Background Information‬

‭Pre-existing Applications‬

‭The primary method/technology that attempts to address AI powered help desk‬ ‭automation is agentic workflows, which are composed of large language models (LLMs),‬ ‭application programming interfaces (APIs) and automation frameworks (Shrivastava, 2025) In‬ ‭this instance, an agent assigns a priority label (low, medium, high) to incoming help desk‬ ‭requests based on the user’s input. If the request is high priority, it is immediately escalated to a‬ ‭human agent. Otherwise, the agent determines whether the query can be resolved autonomously‬ ‭using an LLM that leverages retrieval-augmented generation (RAG). If the agent is still unable to‬ ‭resolve the issue after executing all of the aforementioned steps, then the request is ultimately‬ ‭escalated to a human agent.‬ ‭While the agentic approach is widely used in the industry, we chose to leverage a‬ ‭bidirectional LSTM-based deep learning model for text classification . Similar to the transformer‬ ‭architecture, this approach processes the input sequence in both forward (left to right) and‬ ‭backward (right to left) directions, which allows us to capture contextual information from both‬ ‭past and future tokens (GeeksforGeeks, 2025). This improves the model’s ability to analyze text‬ ‭more accurately, which in turn enhances priority tagging in our help desk automation workflow.‬ ‭Our Approach‬

‭11‬

59