III · The Observatoryتفكّر
Research
Papers written from systems that run in production, with the code and data released so every number in them can be checked.
Yahya Elnawasany
AI Engineer · Research & Deployment
B.Sc. Artificial Intelligence, Kafr El-Sheikh University, 2026
Yahya Elnawasany works on Arabic language and speech: Quranic text processing, speech recognition and synthesis, and language models that answer from verified sources instead of from memory.
Most of his research starts as a problem inside a system people actually use. The validator in his first paper was built for «مُسلم», the voice AI platform he runs in production, and was measured there before it was written up.
He releases the code and data with each paper, so the numbers can be checked rather than trusted.
Papers
Sole-authored. Newest first.
- arXiv:2609.31511Systems papercs.CLSeptember 2026
Muslim: A Deployed Arabic Voice AI Platform for Grounded Islamic Knowledge
The production platform behind the validator: a real-time Arabic voice pipeline, a released model family, and the metering and observability that let strangers use it.
- arXiv:2609.14967Resource papercs.CLSeptember 2026
A Corpus-Aligned Uthmani-to-Standard Quranic Word Mapping and a Deterministic Recitation Validator
A 2,290-pair word mapping that closes the gap between the Quran’s Uthmani script and the Standard Arabic that NLP tools are built for, and the deterministic recitation validator built on it.
Released with the papers
Open code, data and models, so the claims can be checked and the work reused.
- ModelMuslim-6B-PROThe tool-routing language model that answers users in production.Muslim: A Deployed Arabic Voice AI Platform
- ModelMuslim-6B-PRO-GGUFQuantized build of the same model for local inference.Muslim: A Deployed Arabic Voice AI Platform
- ModelFasih-TTS-V1The self-hosted Modern Standard Arabic voice.Muslim: A Deployed Arabic Voice AI Platform
- DatasetFasih-TTS-BenchmarkEvaluation set for the voice.Muslim: A Deployed Arabic Voice AI Platform
- Code & datamuslim-quran-validatorThe mapping, the script that rebuilds it, the validator, and the 124-case evaluation harness.Uthmani-to-Standard Quranic Word Mapping
What the work is about
- Arabic and Quranic text processing
- Orthography, normalization and alignment between the Uthmani script and Standard Arabic, and what general-purpose tools get wrong about both.
- Arabic speech
- Recognition and synthesis for Modern Standard and Quranic Arabic: fine-tuning, diacritization, and evaluation that survives contact with real users.
- Grounded language models
- Tool-calling models that retrieve from verified sources instead of recalling them, where a misattributed citation is a real harm.
- Systems that stay up
- What a research prototype needs before strangers can use it: metering, capacity, observability, and honest measurement.
Correspondence
Questions about a paper, a result you could not reproduce, or an idea for working together: yahya@yahyaelnawasany.com