Skip to content

AbhiramVSA/bantr-presentation

Repository files navigation

Bantr — AI-Powered Real-Time Voice Debate Platform

An interactive research presentation exploring the architecture, AI pipeline, and infrastructure behind Bantr, a platform that enables real-time voice debates between humans and configurable AI agents.

Built as an academic showcase for Applications of Artificial Intelligence, covering NLP, deep learning, speech processing, RAG systems, and production-grade deployment.

Overview

Bantr pairs users with AI debate opponents that have distinct personalities and argumentation styles. The core pipeline — Speech-to-Text → LLM Reasoning → Text-to-Speech — achieves sub-2 second end-to-end latency. Post-debate analysis includes fallacy detection, argument scoring, sentiment tracking, and mood-based image generation.

This presentation walks through the full system across 11 interactive slides, each backed by looping HLS video and a glassmorphic UI.

Presentation Topics

Slide Content
Cover Project title, team, animated background
Abstract System overview, real-time voice pipeline, post-debate intelligence
Area & Domain Five interdisciplinary AI domains explored
ML Concepts STT → LLM → TTS pipeline breakdown
Deep Learning Transformer architecture (DeBERTa-v3), generative models (DALL·E 3), neural TTS
NLP Speech recognition, language generation, sentiment analysis
RAG & Memory Chunking, vector embeddings, Pinecone retrieval, cross-session persistence
Software & Platforms Frontend, backend, AI services, DevOps stack
Hardware & Infrastructure GPU compute, cloud services, CDN, monitoring
References Academic papers across LLMs, speech, RAG, NLP, and image generation

Tech Stack

Presentation

  • React 19 + TypeScript
  • Vite 7
  • Tailwind CSS 4
  • HLS.js for video streaming
  • Lucide React for icons

Bantr Platform (Presented)

  • Speech: Deepgram Nova-2 (STT), ElevenLabs Turbo v2 (TTS)
  • Reasoning: GPT-4 Turbo, Claude 3.5 Sonnet
  • Classification: Fine-tuned DeBERTa-v3 (14 fallacy types, 89% F1)
  • Retrieval: OpenAI embeddings + Pinecone vector DB
  • Generation: DALL·E 3 for mood-based imagery
  • Backend: Node.js/Express, Python FastAPI, WebSocket, WebRTC
  • Data: PostgreSQL + Prisma, Redis
  • Infra: Vercel, Railway, AWS (S3, CloudFront, g5.xlarge GPU), GitHub Actions

Getting Started

# Install dependencies
npm install

# Start development server
npm run dev

# Build for production
npm run build

# Preview production build
npm run preview

Navigation

  • Arrow keys or Spacebar — navigate between slides
  • F — toggle fullscreen
  • Mouse — bottom controls with progress indicator

Team

  • Abhiram
  • Ishita
  • Pavan

License

This project is part of an academic research presentation and is not licensed for commercial use.

About

Interactive research presentation for Bantr, an AI-powered real-time voice debate platform built with React and Vite.

Topics

Resources

Stars

Watchers

Forks

Contributors

Languages