Project: Talking kAIoska

Team Name - Talking kAIoska

Problem Statement - GenAI-Powered Customer Support Process Optimization

Team Leader Email - uralstech@gmail.com

A Brief of the Prototype:

kAIoska: AI Talking Multilingual No-Touch Avatar-Based Information Kiosk Customer Service agents

Benchmark:

• IPEX Optimized average inference time (sec): 1.2261635462443035
• Not optimized average inference time (sec): 1.812053879102071

Tech Stack:

• Server:
  • Python-powered
  • Running on Intel Developer Cloud
  • Using HuggingFace Transformers w/ acceleration from Intel Extensions for Pytorch (IPEX), utilizing OneAPI
  • TinyLLama as the model - Quantized and optimized by IPEX
  • FastAPI for simple REST API
  • Authentication using Firebase Admin SDK
• App:
  • Powered by Unity w/ C#
  • ARCore/OpenCV for face tracking
  • Whisper AI for spoken language detection
  • Google cloud STT, TTS, Translate for speech-to-text-to-speech processing
  • Server app for main LLM

Step-by-Step Code Execution Instructions (Server):

• Create a Firebase project
• Create a service account on connected GCP project
• Download json key file
• Run:

export GOOGLE_APPLICATION_CREDENTIALS="path/to/service/account/key/"

pip install requirements.txt
python src/main.py

Future Scope:

• Integration with CRMs
• Call answering/forwarding
• AI Training & Hyper-personalization