ScreenMan

Transcript view
Stage view of related context
View of recommended topic transitions

Inspiration

Inspired by The Joe Budden podcast's production approach, where a behind-the-scenes team member researches topics in real-time to enrich discussions. As a podcast host, I recognized how valuable this role is for maintaining engaging, fact-based conversations. ScreenMan automates this support function, allowing hosts to focus on the conversation while maintaining accuracy and flow.

This could further be applied to live streamers and live audio hosts on sports and radio shows.

What it does

ScreenMan is a web application that serves as an AI-powered research assistant for live conversations. Its key features include:

Real-time speech transcription and context analysis
Jeopardy-style display board showing relevant facts and media
Topic transition suggestions

How we built it

Built on a modern tech stack combining real-time audio processing with AI services:

Python server for audio buffering and OpenAI Whisper integration
NextJS full-stack application managing the frontend and backend pipelines
Multi-stage AI pipeline for:
- Speech transcription
- Claim detection and search query optimization
- Web research and fact verification
- Context-aware content staging
ElevenLabs text-to-speech for prompting narration
PostHog for analytics and LLM generation tracking

Challenges we ran into

Developed a text simulation system to accelerate testing and reduce API costs
Engineered prompts for consistent, structured outputs across each distinct AI task

Accomplishments that we're proud of

Created an intuitive interface that masks complex AI operations
Achieved near real-time performance for the research pipeline (15secs vs a human 1min+)
Building a practical tool that solves a real problem in content creation