This talk presents Not For Her, an interactive installation created for the Inequalities exhibition at Triennale Milano. The goal was to design an immersive experience that makes visitors feel the dynamics of gender discrimination—not through data or statistics, but through a simulated job interview with virtual avatars.

From a technical perspective, the challenge was to orchestrate a complex system integrating:

• Controlled Generative AI: An LLM (GPT-4.1) guided by a state machine and dynamic prompts to ensure narrative coherence while avoiding incoherent or trivial responses.
• Speech Recognition and Synthesis: An optimized audio pipeline with local Whisper STT and streaming TTS to minimize latency and handle natural interruptions.
• Emotion Analysis: A ResNet-based model for real-time facial emotion recognition, feeding into the prompt to modulate tone and content.
• 3D Avatars: Lip-sync based on phonemes and dynamic animations for credible interaction.
• Privacy and Compliance: Local edge computing with no biometric data persistence, fully GDPR-compliant.

The talk will cover:

• System Architecture: How Python served as the backbone to orchestrate AI modules, state machines, and audio/video pipelines.
• Advanced Prompt Engineering: Techniques to balance creativity and control, with examples of how prompts adapted to user context and emotional state.
• Technical Challenges and Solutions: Latency reduction, interruption handling, and robustness in high-traffic environments.
• Impact and Evaluation: Qualitative and quantitative results (over 3,100 individual experiences, international audience), confirming the effectiveness in raising awareness.

This project demonstrates how generative AI can be channeled into structured narratives, transforming from a simple text engine into a tool for meaningful experiences. The talk targets developers, data scientists, and creatives interested in building interactive systems with Python and AI, focusing on real-time architectures and narrative control.