Track: VoiceTech |
| INSIGHTSTACK: Scaling Insights Extraction from Customer Feedback |
| Organizations today collect an overwhelming volume of unstructured customer feedback from sources which contains valuable insights about customer sentiments, recurring issues, and service gaps that are crucial for business success. Manually sifting through is impractical at scale. While, large language models (LLMs) offer a way to summarize and extract themes from text, they face limitations: LLMs cannot process thousands of feedback entries at once (due to input size constraints), they also sometimes overgeneralize or hallucinate trends, and it’s hard to trace which specific comments lead to a given AI-generated insight. Our project, Insight Stack, addresses these challenges by combining the strengths of clustering algorithms with the power of LLMs to analyze customer feedback at scale. We use clustering and topic modeling to automatically group similar customer comments into coherent themes (making the insights traceable and trustworthy). We then apply LLM-driven analysis within each cluster to produce a concise, human-readable summary of that theme – complete with the prevailing sentiment and key issues highlighted. My presentation will compare embedding choices, topic generation strategies, and data input variations, showing how each setup affects coherence, sentiment accuracy and demonstrating the human validated approach which is most effective for our use case. The goal is to give product and support teams a system that surfaces what to fix, why it matters to customers, and how issues evolve over time, so they can act faster, reduce churn, and allocate resources where they deliver the greatest impact. |
|
| Presentation Video |
| Presentation Notes |
|
Acquaye-INSIGHTSTACK.pdf |
Real Time Communications Conference & Expo at Illinois Tech
IEEE International Conference
Follow Us
Share This
