Services · Arabic AI
Most LLM tools are built and tested in English. Arabic brings its own problems: Modern Standard Arabic next to Gulf, Egyptian and Levantine dialects, Arabic and English mixed in one sentence, several valid spellings of the same word, and right-to-left documents that confuse standard parsers. A system that scores well in English can quietly fail your Arabic-speaking users.
I'm a native Arabic speaker and a Senior AI Engineer in Riyadh. I build Arabic and bilingual assistants, search and speech pipelines, and I test them on Arabic the way your users actually write and speak it.
The same word can be written several ways: different alef and hamza forms (أ إ آ ا), taa marbuta or haa (ة / ه), alef maqsura or yaa (ى / ي), with or without diacritics. People searching for my own name type both إسلام and اسلام. Without normalization, search misses obvious matches.
Users write in dialect, switch to English for technical terms, or type Arabic in Latin letters. Retrieval and prompts have to handle all of it, and the evaluation has to include it.
Many models split Arabic text into more tokens than the same content in English. That raises cost and latency and reduces how much fits in the context window, so the choice of model and embedding model matters more than it does in English.
Scanned PDFs, mixed-direction tables and OCR noise need cleanup before indexing. For audio, Whisper large-v3 transcribes Arabic, but accuracy varies with dialect and domain vocabulary, so transcripts have to be checked against real recordings.
Question answering over Arabic and English documents, answering in the user's language and citing the source passages.
Search that matches meaning, not just exact words, with spelling normalization and hybrid keyword search for names, codes and numbers.
Whisper transcription of calls, meetings and recordings, feeding search, summaries, information extraction and analytics.
LoRA and QLoRA fine-tuning of open-weight models for Arabic classification, entity extraction and domain-specific answers, when prompting alone isn't enough.
Real questions and documents from your domain, in Modern Standard Arabic and dialect, including mixed Arabic–English text. Not English questions run through a translator.
Open-weight multilingual models such as Qwen, LLaMA and Mistral when data must stay on your servers; API models such as GPT, Claude and Gemini when it's allowed to leave; and Arabic-first models such as ALLaM or Jais when they fit. Embedding models are compared the same way.
Text normalization, chunking, hybrid search and re-ranking, each measured on the Arabic test set rather than assumed to carry over from English.
Deployment on your cloud or your own servers, with feedback from Arabic-speaking users feeding the next round of tests.
RAG assistants, LLM chatbots and Whisper large-v3 speech-to-text for large-scale Saudi enterprise use cases, running on self-hosted LLaMA and Qwen models on GPU Kubernetes.
Impact: data-sovereign generative AI for privacy-critical environments.
LLM, RAG and Whisper systems that automate recruitment and workforce workflows, including semantic resume–role matching and an AI meeting assistant.
Impact: deployed in government hiring workflows.
Fine-tuned NLP models for classification, named-entity recognition and semantic search for clients worldwide, delivered with a 100% Job Success Score on Upwork.
Both. The test set includes the dialects your users actually use, so you see accuracy for each one before launch.
Yes. Open-weight models, embeddings and the vector database can all run on your infrastructure or in a cloud region inside Saudi Arabia.
It depends on your data, your budget and where the system has to run. I don't pick one in advance: I compare a short list on your Arabic test set and show you the results.
Usually with a short consultation and a small proof of concept on your own Arabic documents or recordings.
Tell me what your users need and how they write. I reply in Arabic or English.