AI Smart GID

AI Smart GID is a voice assistant for a museum information kiosk. It detects whether a visitor speaks Uzbek, Russian or English and answers by voice through a talking avatar, grounded in the museum's knowledge base.

Live: aigid.uzavtoyul.uz

A voice guide in a museum has to understand three languages, cope with names that speech recognition misspells, and never invent facts. When it does not know something, it should say so.

I built the whole product on my own:

• the full voice pipeline: speech-to-text, language detection, moderation, RAG, an LLM answer (Gemini or GPT) and text-to-speech (Azure TTS);
• an admin panel with roles, statistics and charts, CSV export, knowledge-base editing and Word/PDF upload with automatic indexing;
• production deployment on Linux (systemd, Nginx, HTTPS, rate limiting) with GitLab CI/CD, and the Windows 11 kiosk setup.

• Anti-hallucination flow. RAG first, then a broadened retry with a background web search. Facts are verified against trusted sources; otherwise the guide replies "no information" instead of guessing.
• Hybrid retrieval. Vector search (FAISS), keyword search (BM25) and trigram similarity together, so names misspelled by speech recognition are still found. Every change is measured on an evaluation set (hit@k, MRR).
• Web search. Parallel search (Tavily) over trusted sites and the open web, with date-aware ranking and verification of numbers, dates and names.
• Live avatar. A streaming avatar (Anam, WebRTC) that visitors can interrupt.

• The median number of sources behind an answer grew from 2 to 8.
• Billed avatar time was cut roughly threefold.