Najot Nur
NotiqAI
An AI speech coach for Uzbek: letter-by-letter pronunciation analysis inside a course, audiobook and certification platform.
- Client
- Najot Nur
- Role
- Lead engineer: backend, AI core, mobile and web
- Period
- Jun 2026 – Sept 2026
- Stack
- Python 3.12FastAPISQLAlchemyPostgreSQLRedisWhisperGroqGeminiClaudeFlutterReactViteTailwindPaymeDockerCaddyGrafana
Problem
No pronunciation analysis exists for Uzbek; a coach had to listen to every learner live.
Solution
A deterministic phoneme-alignment core + an LLM coaching layer + a full course and payment platform.
Result
notiqlik.uz is live: a mobile app, four web apps, 27 migrations, UI in three languages.
The Najot Nur public-speaking centre in Tashkent wanted to scale speech and pronunciation feedback beyond live coaching. The problem: speech AI for Uzbek phonology (k/q, g/gʻ, o/oʻ, sh/s, ch/c, ng/n) barely exists.
The solution has two layers. The core is deterministic Python: Levenshtein alignment, Uzbek grapheme tokenisation (oʻ, gʻ, sh, ch, ng) and per-letter penalties. The learner reads a reference text, the system scores every word and marks wrong sounds in red. LLMs only add coaching prose on top (tongue position, minimal pairs), so the app keeps working even when an AI provider is down. The speech check finds filler words, pauses and clarity in a two-minute talk.
The platform is complete: video lessons and quizzes, curator grading, a three-stage content workflow, audiobooks, Payme/Atmos/Uzum Nasiya payments and PDF certificates. Content protection: a moving watermark, screenshot detection and server-side sessions.
Stack: FastAPI, async SQLAlchemy, Redis, Groq Whisper large-v3, a multi-provider LLM layer (Groq Llama 3.3, Gemini 2.5, Claude) with a parallel ensemble, a Flutter app, four React apps, Docker + Caddy, Prometheus/Grafana/Loki, GitHub Actions → ghcr.io + Watchtower.