AAI_2025_Capstone_Chronicles_Combined

‭ResolveAI‬

‭One of the unexpected observations was the misclassification of some tickets that seemed‬ ‭to request human interaction or use terms to magnify the issue, even though the issue itself was‬ ‭low priority. This indicates that user intent may sometimes override true technical urgency,‬ ‭which highlights the nuance and subjectivity in labeling ticket priority. The experiment used‬ ‭semi-generic text features and we believe that introducing domain specific text features/informed‬ ‭heuristics can help alleviate this issue.‬ ‭If we were to continue this project, the next steps would focus on improving the model’s‬ ‭performance and preparing it for real-world deployment. A key priority would be to work with a‬ ‭more defined and domain-specific dataset, as the current semi-generic data likely contributed to‬ ‭misclassifications, particularly around priority levels. To address data imbalance and skew, we‬ ‭could explore synthetic data generation techniques using tools like InstructLab to enrich‬ ‭underrepresented classes.‬ ‭Additionally, integrating domain-specific text features and informed heuristics such as‬ ‭keyword patterns or urgency indicators could enhance the model's understanding of ticket intent‬ ‭and priority. Another step toward productionizing ResolveAI would be building integrations with‬ ‭widely used help desk platforms like Jira or ServiceNow to allow seamless routing of tickets‬ ‭based on model predictions.‬ ‭Given the increasing relevance of AI agents and agentic workflows, we also envision‬ ‭transitioning this tool to be part of a larger, intelligent support ecosystem that dynamically‬ ‭handles and escalates tickets using LLM-powered reasoning components alongside traditional‬ ‭models.‬ ‭Next Steps‬

‭23‬

71