BKSBhupender Kumar Saini
Research focusCV

Oct 2025 – Dec 2025/Industry Project

Conversational Analytics Platform

  • Industry Project
  • Industry Project with Automotive Company
Partner
Automotive Company
Period
Oct 2025 – Dec 2025
Focus
Industry Project

Key outcomeIntegrated prototype covering the interface, backend agents, API integration, visualization, and evaluation logic

Conversational analytics dashboard showing conversation volume over time, emotion breakdown, service type, time of day, and intent counts such as navigation, climate, and media
Dashboard overview: conversation volume, emotion, service type, and intent breakdown.
Conversation analysis detail view showing a driver-assistant dialogue with sentiment tags and driver experience, interaction quality, and intelligence scoring
Per-conversation analysis: sentiment, driver experience, and interaction-quality scoring for each exchange.

Context

Conversational systems generate far more qualitative feedback than any team can read by hand. This project explored how to turn that raw conversation data into structured, explainable signals for product and UX teams, without hiding the reasoning behind each recommendation.

Inputs

  • Conversational feedback and survey responses
  • Behavioral and contextual signals

Approach

  • Designed the interaction flow and built frontend components in TypeScript and SvelteKit, including analytical and comparison visualizations
  • Built Python data-processing pipelines and a multi-agent architecture on the OpenAI Agent SDK, with agents for classification, tagging, and contextual evaluation
  • Developed an aggregation agent that turns individual agent outputs into ranked, explainable recommendations
  • Evaluated AI-generated recommendations against human expert judgments

Outcomes

  • Integrated prototype covering the interface, backend agents, API integration, visualization, and evaluation logic
  • A traceable route from raw conversation to classification, interpretation, and final recommendation
  • Recommendations become more useful when that route stays visible and testable, rather than arriving as a single opaque score

Related on the shelf