AI Voice Assistant for Mobile Operator

## About the Project

Intelligent voice and text assistant for Ucell mobile operator (Uzbekistan), capable of consulting clients on tariff plans, answering frequently asked questions, and providing personalized recommendations based on user needs.

## Challenges Solved

- **Consultation Automation**: Reducing call center load through automatic responses to common questions
- **Personalization**: Smart tariff selection based on user needs analysis (internet, calls, SMS)
- **Multilingualism**: Full support for Russian and Uzbek languages
- **24/7 Availability**: Round-the-clock operation without downtime

## Key Features

**Voice and Text Interface**
- Speech recognition and synthesis via Yandex SpeechKit with native voices for RU/UZ
- Support for WebM, Opus, MP3 formats
- Text chat for written communication

**Intelligent Recommendation System**
- NLP analysis of user requirements
- Vector search across 50+ tariff plans database
- Personalized recommendations based on needs

**FAQ with Semantic Search**
- Knowledge base: 29 Q&A in 13 categories
- Vector search with 87-98% accuracy
- Automatic vectorization of new FAQ
- View statistics for popularity analysis

**Advanced Admin Panel**
- Tariff and FAQ management through user-friendly interface
- Inline editing, question similarity testing
- Detailed dialogue logs with time metrics
- Request statistics visualization

## Technology Stack

**Backend**: Django 5.2 (async), Django Ninja (REST API), PostgreSQL 16 + pgvector, Redis

**AI & ML**: OpenAI GPT-4, Yandex SpeechKit (STT/TTS), sentence-transformers (multilingual-e5-large), pgvector (vector search)

**DevOps**: Docker & Docker Compose, Gunicorn + Uvicorn, Nginx, Systemd

**Additional**: django-unfold, FFmpeg, cryptography, httpx

## Technical Features

**Asynchronous Processing**: Parallel STT, vector search, AI generation, and TTS work to minimize response time

**Vector Search**: Semantic comparison with 0.7 threshold for FAQ, vector caching for acceleration

**Contextual Dialogues**: Storing last 10 messages history, continuous dialogues with session_id, adaptive prompts

**Analytics**: Time metrics for each stage, token counting, full request logging

## Results

- Request processing: < 2 seconds for full cycle (STT → AI → TTS)
- FAQ accuracy: 87-98% semantic search relevance
- Coverage: 29 FAQ in 13 categories, 50+ tariff plans
- Security: API tokens, data encryption (Fernet), CORS/CSRF protection, rate limiting
- Production-ready: Docker containers, automatic migrations, health checks, SSL/TLS

## Achievements

The project demonstrates deep understanding of modern AI/ML technologies, experience with vector databases, skills in creating high-load async systems and integrating complex external APIs (Yandex, OpenAI), knowledge of DevOps practices.

---

**Technologies**: Python, Django 5.2, PostgreSQL, pgvector, Redis, Docker, OpenAI GPT-4, Yandex SpeechKit, NLP, Vector Search, REST API, Async/Await

**Development Time**: 3 weeks | **Status**: Production-ready, actively used