PORTFOLIO

Developed an advanced IT support ticket classification system using a Large Language Model (LLM) to automatically categorize tickets, assign relevant tags, set priority, estimate resolution time, and draft an initial response. Leveraged the Mistral-7B-Instruct model with prompt engineering and few-shot learning to achieve sophisticated, structured outputs with a small dataset, showcasing a modern approach to NLP-driven automation in enterprise support systems.

Support teams often waste valuable time manually categorizing tickets, leading to slower resolution times and inconsistent prioritization. The goal was to design a proof-of-concept NLP system that could reliably interpret and process support ticket text to streamline the entire triage process.

The solution was built iteratively, increasing the complexity and capability of the LLM at each stage through carefully engineered prompts.

• Utilized a sample dataset of 21 IT support tickets.
• Loaded the Mistral-7B-Instruct-v0.2-GGUF model locally using the llama_cpp library, requiring no model training or fine-tuning.

• Prompt 1 (Categorization): Instructed the model to classify each ticket into "Hardware," "Software," or "Network" and return the output in a structured JSON format.
• Prompt 2 (Tag Generation): Enhanced the prompt to also extract the most relevant keyword as a "Tag" for easier search and trend analysis.
• Prompt 3 (Priority & ETA): Further refined the prompt to have the model assign a "Priority" level and an estimated time of arrival ("ETA") based on the ticket's urgency.
• Prompt 4 (Response Generation): The final prompt directed the model to generate a complete JSON object including all previous fields plus a helpful, context-aware "IT Response Message".

• The model successfully processed all tickets and generated structured JSON outputs for each task, demonstrating the effectiveness of the few-shot prompting technique.
• Analysis of the outputs showed a logical distribution of categories, with Hardware issues being the most frequent in the sample data.
• The system correctly identified urgent tickets, assigning High priority and Immediate ETAs to the majority of requests.

• Faster Ticket Triage: Immediate categorization and prioritization save valuable time for support agents.
• Consistency: Reduces human variability in categorization and ensures a standardized initial response to users.
• Scalability: This prompt-based approach is highly adaptable and can be scaled to handle thousands of tickets and more granular categories without retraining a model.

• Expand the dataset to further validate the model's performance across a wider range of issues.
• Integrate the system into a live ticketing platform via an API to automate the workflow in real-time.
• Develop a formal evaluation framework to measure the accuracy of the model's classifications and the quality of its generated responses.