PULSE: A Real-Time LLM-Powered Sentiment Intelligence Platform
Ankit Kumar¹, Archita Bal², Shubhakanta Behera³, Soumik Routray⁴, B Ujalesh Subudhi⁵
¹ B.Tech Student, Department of Computer Science and Engineering, NIST University, Berhampur, India
² B.Tech Student, Department of Electronics and Communication Engineering, NIST University, Berhampur, India
³ B.Tech Student, Department of Computer Science and Technology, NIST University, Berhampur, India
⁴ B.Tech Student, Department of Computer Science and Technology, NIST University, Berhampur,
India
⁵ Assistant Professor, Department of Computer Science and Engineering, NIST University, Berhampur,
India
Abstract
PULSE is a real-time sentiment intelligence platform that is distributed where large language models (LLMs) are used. have been used to analyze user-generated text multi-dimensionally in terms of affect. Text input is welcomed. via a REST interface or Web interface and categorized in six Ekman emotion dimensions, generating. confidence-scored sentiment labels, sentence-level heatmaps, and extractions of keyword polarity. With a median latency of less than 300 ms, Google Gemini 1.5 Flash returns structured JSON responses. configured as the default inference server, and automatic failover to Groq LLaMA 3.1, and a 280x Redis SHA-256 content-hash caching is used to reduce latency. The results are saved in a. multi-tenant PostgreSQL schema. As an optional addition, a Behavioral Context Injection (BCI) it includes the module which observed browser-native keystroke dynamics - including typing speed. correction rate, and pause distribution — and adds an emotional state label to the prompt to the LLM as derived. as supplementary context. This is augmented to live interactive interfaces where typing occurs. behaviour is a helpful cue to disambiguating text which is emotionally ambiguous. Evaluation on the IMDb dataset (50,000 reviews) and the SST-2 benchmark (67,349 sentences) yields 94.2% and 92.8% binary sentiment accuracy respectively, and a 3.2-percentage-point improvement seen on ambiguous inputs. when BCI module is switched on.
Keywords: Sentiment Analysis, Large Language Models, Real-Time Systems, Distributed Architecture, Affective Computing, Redis Caching, Multi-Tenant API, Behavioral Context Injection, IMDb, SST-2
