Amir Bawab

Saudi Arabia and NVIDIA will build AI factories powering the next wave of intelligence. NVIDIA will help deploy AI infrastructure, strengthen the regional ecosystem, and train developers and scientists – fueling growth and prosperity in the region. Read more.

7

What if code could decorate itself? 🎨 This is Bayan (بيان) — an Arabic programming language. In this video, a Bayan program draws a frame around its own code, stroke by stroke. Self-decorating code. In Arabic. In the browser. Built for Arabic-speaking students who deserve to learn programming in their own language — not translated, not adapted. Native. 👉 bayan.ink #ArabicTech #EdTech #CreativeCoding #ProgrammingLanguages #CSEducation

6

1 Comment

Saudi Arabia and NVIDIA will build AI factories powering the next wave of intelligence. NVIDIA will help deploy AI infrastructure, strengthen the regional ecosystem, and train developers and scientists – fueling growth and prosperity in the region. Read more.

4

100 million Egyptians don't speak the Arabic that AI understands. Most Arabic NLP models train on Modern Standard Arabic—a formal language nobody speaks in daily conversation. So I built something to help fix this: 4,322 Egyptian Arabic dialogue pairs, fully open-source. - Natural conversational language - 18 domain categories - Ready to use for translation, chatbots, research Dataset: https://lnkd.in/g6HE3kvN Thoughts: https://lnkd.in/g_McQdib #NLP #MachineLearning #Arabic #OpenSource

55

8 Comments

this is actually insane (in the best way) Elijah Ghossein and Ali Hamad saw how research collaboration is and instead of complaining, they built something better. no code. no gatekeeping. just execution. that’s the kind of energy that stands out. lowkey the biggest takeaway: → you don’t need perfect tools, you need a clear problem → builders > observers → the barrier to creating real impact is basically gone decoded brain feels like one of those early signals 👀 Notion

1

🗣️ Habibi, Speak: The Rise of Dialectal Arabic AI Voices How I’m Building and Expanding Habibi-TTS Locally — and Why It Matters ------------------------ 🌍 The Problem: AI Doesn’t Speak Like Us Most speech synthesis systems today are trained on English — and even when they support Arabic, they often default to Modern Standard Arabic (MSA). That’s useful for formal settings, but it’s not how people actually speak. From Cairo to Casablanca, Riyadh to Ramallah, Arabic lives in its dialects. It’s emotional, musical, and deeply local. And until recently, AI couldn’t speak it — not fluently, not naturally, and certainly not with cultural nuance. ------------------------ 💡 The Breakthrough: Habibi-TTS That’s why I started working with Habibi-TTS — an open-source project that lays the foundation for unified dialectal Arabic speech synthesis. It’s built on cutting-edge models and trained on diverse datasets from Hugging Face, covering multiple dialects and phonetic styles. I’ve successfully deployed Habibi-TTS locally on CPU, and it’s running smoothly. No GPU required. This makes it accessible for developers, researchers, and educators across the Arab world — even those without high-end hardware. ------------------------ 🧪 What I’m Doing Now I’m now expanding Habibi-TTS into new dialects, starting with Gulf Arabic and Egyptian colloquial. I’m experimenting with: - Fine-tuning phoneme alignment for dialectal pronunciation - Custom voice cloning using local speakers - Integrating SILMA TTS APIs for real-time synthesis and comparison - Exploring Kaggle pipelines like Arabic_Dialect_TTS_SAU for Saudi dialect modeling This isn’t just technical work. It’s cultural preservation. It’s about giving voice to communities that have been digitally silent for too long. ------------------------ 🔍 Why This Matters Arabic is one of the most spoken languages in the world — but it’s also one of the most linguistically diverse. Dialectal TTS unlocks: - Education: Teaching children in their native dialect - Accessibility: Helping visually impaired users navigate tech in familiar speech - Storytelling: Enabling podcasts, audiobooks, and games with authentic voices - Healthcare: Supporting patient–AI interactions in local dialects - Cultural AI: Building agents that understand not just language, but identity ---------------------- 🚀 What’s Next I’m planning to: - Release sample voices across dialects - Benchmark CPU performance across models - Contribute to open datasets and phoneme maps - Collaborate with researchers and developers across the region If you’re working on Arabic NLP, speech synthesis, or cultural AI — let’s connect. This is a movement, not just a model. --------------------- 📢 Final Thought AI should speak like us. Not just in our language — but in our voice. Habibi-TTS is a step toward that future. And I’m proud to be building it, one dialect at a time. #HabibiTTS #ArabicAI #VoiceTech #AIandCulture #OpenSourceAI #AI2026 #SILMA #HuggingFace

15

4 Comments

AI models speak fluent English. But ask them to handle Arabic (or Thai, Telugu, etc.) — and they struggle. Billions of people are left out. Why? Because teaching AI new languages is expensive and risky. (Break the bank or break the model?) Researchers just found a better way - and dropped "Kuwain 1.5B" They taught a tiny English model to speak Arabic without: ➤ Retraining everything from scratch ➤ Ruining its English skills ➤ Burning 7 figures on GPUs 𝐓𝐡𝐞 𝐫𝐞𝐬𝐮𝐥𝐭? ➤ 70% cheaper to build bilingual models ➤ English stays sharp, Arabic levels up ➤ A blueprint for making AI truly global 𝐖𝐡𝐲 𝐬𝐡𝐨𝐮𝐥𝐝 𝐲𝐨𝐮 𝐜𝐚𝐫𝐞? Because this technique could make multilingual AI accessible to everyone — startups, indie hackers, open-source devs. Sandra & I cover the full story in today’s edition of The Vision Debugged - AI Newsletter. 🔗 https://lnkd.in/g7n2BDV4 Found this valuable? Repost to help more builders discover it. ♻️ ______________________________ PS. Follow me (Tezan Sahu) for more AI breakthroughs. PPS. Subscribe to the newsletter for more such weekly deep dives - https://lnkd.in/gwUNCBdJ

21

1 Comment

🚀 Breaking: Egypt launches Karnak, the largest Arabic language model to date! 🇪🇬 At AI Everything Summit in Cairo (Feb 2026), ITIDA unveiled Karnak – a massive LLM with 30-40B and 70-80B parameters. Trained to master Arabic dialects, local contexts, and cultural nuances, it powers apps like "Sia" for education and legal tools for SMEs. This national AI powerhouse boosts digital sovereignty, empowers startups, and reduces reliance on foreign models like ChatGPT. A game-changer for MENA tech! #KarnakAI #EgyptAI #ArabicLLM

38

3 Comments

If Future, async, and await ever confused you in Dart… don’t worry That’s why I made this video for you, to explain: ✅ What they mean ✅ Why we use them ✅ How they work with simple, real examples Feedback is always welcome 🤗❤️ #Dart #Flutter #AsyncAwait #ArabicContent #SoftwareEngineering #LearningInPublic

65

4 Comments

Stay tuned for tomorrow's episode of Pioneers of AI where Rana el Kaliouby, Ph.D. sits down with Sal Khan, founder of Khan Academy for an energizing conversation on how to bring AI into the classroom ethically and responsibly.   Don't miss this timely back-to-school episode!

92

6 Comments

Big news from the UAE: Free ChatGPT Plus access for all residents, a bold move that puts AI in everyone’s hands. This isn’t just about tech, it’s about vision, access, and opportunity. Just Amazing!

5

The future of programming is hybrid. Our Chief AI Officer, Dr. Rania Khalaf, shares her insights on Forbes, explaining why the ability to blend code and natural language is the key to innovation. Read the full article to learn more! https://lnkd.in/gVVYd5g2

16

$15.2B is not an announcement. It is a foundation pour. Microsoft is not “partnering” with the UAE for optics. It is embedding itself into national infrastructure. Data. Compute. Governance. This is what real AI investment looks like. Long horizon. Sovereign alignment. Physical and institutional build-out. Notice what is missing. No talk of classroom tools. No pilot programs. No dashboards. Because this layer sits below all of that. For education leaders, this matters more than it appears. Curriculum debates happen at the surface. Power is decided at the substrate. Compute location. Data governance. Regulatory alignment. Those are load-bearing decisions. If your AI strategy assumes these layers are neutral or interchangeable, that is a structural flaw. You are designing classrooms without knowing who owns the steel. AI is not changing education yet. But it is quietly changing the ground education will be built on. Most systems are not looking down far enough. https://lnkd.in/gHqkuzci.

2

Excited to share the recording of my session at QatAIyef: AI Engineering Nights, organized by IEEE EUI Computer Society SBC . 🎤 From Repo to Production: Deploying AI Apps the Right Way We discussed how deploying AI systems goes beyond just shipping code, covering architecture decisions, model hosting options, governance, observability, and production readiness. 📺 Session recording: https://lnkd.in/dHz8ijF6 #AI #GenerativeAI #LLM #SoftwareEngineering #DevOps #Cloud #PlatformEngineering #IEEE

45

Sub-10ms Search across 3.4 Million Classical Arabic Segments. Most people think High-Performance AI requires massive GPU clusters and five-figure cloud bills. We just proved them wrong. I’m excited to share a project I’ve been building: Bahtsul Masail (bahtsulmasail.tech). It’s a Retrieval Augmented Generation (RAG) system designed specifically for classical Islamic jurisprudence. But the real story is under the hood. While most RAG pipelines struggle with latency and cost, we optimized for the "impossible": Search Latency: <10ms (p50). Throughput: 600+ Queries Per Second on a single 8GB machine. Corpus: 3.4 Million text segments from 7,872 classical books. Accuracy: 95% translation accuracy (Indonesian → Classical Arabic). How did we do it? The "Standard AI Stack" (Python/LangChain/Distributed DBs) is great for prototyping, but it often hits a wall at scale. For this project, we took a different path: 1️⃣ The Power of Rust: By using Rust, we achieved zero-cost abstractions and memory efficiency that Python simply can’t match. Our API process consumes just 16MB of RAM at rest. Compare that to the 4GB+ typical of similar Python deployments. 2️⃣ Hybrid Retrieval Cascade: Instead of running heavy semantic searches on every query, we built a tiered system. We use PostgreSQL trigram similarity for precise matches and fallback to Meilisearch and OpenAI embeddings only when necessary. 3️⃣ Domain-Specific Filtering: We didn’t just index text; we indexed tradition. The system is "Madhab-aware," allowing scholars to filter by jurisprudential schools (Shafi'i, Maliki, etc.) with zero performance hit. Why this matters: The Islamic legal tradition (Nahdlatul Ulama) spans centuries of reasoning. Traditionally, finding a specific ruling required years of training and hours of manual book-flipping. We are making this scholarship computationally accessible at interactive speeds. It’s about bridging 1,000-year-old wisdom with sub-10ms modern technology. 🤝 Let’s build together: If you’re looking to build a high-performance system like this—whether for specialized document retrieval, legal tech, or large-scale RAG—let’s talk. I’m open to collaborations and can help you implement this architecture for a highly competitive value. ⚠️ Early Stages: The project is live at https://bahtsulmasail.tech. It’s still in early development and has its flaws, but the foundation is solid. 📖 The Deep Dive: For the engineers and researchers who want to see the math, the benchmarks, and the architectural decisions behind this system, I’ve written a full technical article here: 👉 https://lnkd.in/gPNsnXJZ #AI #Rust #Agent #Claude #RAG #Python

25

2 Comments

From #InformationWarfare to #AlgorithmicWarfare: The Era of "The First AI War” I recently joined Dr. Taghreed Hussein on Nile TV’s "Daily Debate" to discuss a chilling milestone in modern history: the transition of the #MiddleEast conflict into what many now call the "World’s #FirstAIWar." As a specialist in #AIGovernance, I highlighted that we are no longer just using #AI for support; we are seeing AI drive the "kill chain." Systems like "The #Gospel" and #Lavender have transformed the battlefield into a target production factory, shifting the human role from an investigator to a mere "approver." Key takeaways from our discussion: The #Accountability Gap: You cannot court-martial a piece of code. #InternationalLaw is built on #empathy and human judgment, traits a machine can never possess. #FlashEscalation: AI reacts in milliseconds. We are losing the "buffer of time" that 3diplomacy requires, risking accidental global escalation. #AutomationBias: Under the pressure of war, "Meaningful Human Control" is often replaced by a "rubber stamp" on machine-generated scores. We urgently need a #GenevaConvention for AI. #Innovation must never outpace oversight, and the efficiency of the machine must never replace the #morality of the human. My sincere gratitude to Dr. Taghreed for her insightful choice of this critical topic and special thanks to the distinguished program crew for their professionalism 🌹🌹 Link of the episode in the first comment 👇 #DigitalEthics #NileTV #InternationalSecurity #TechLeadership #ISfTeH

69

9 Comments

I just tested GPT-4 on Arabic poetry analysis. It failed spectacularly. Not because the model is bad—but because Arabic isn't just another language. It's 22 dialects, right-to-left script, gender-inflected verbs, and 1,400 years of cultural context that Western AI models were never designed to handle. Here's why your "multilingual" AI is probably broken for 420 million Arabic speakers... 👇 **The Linguistic Challenge** Arabic is a linguistic complexity multiplier: • 22 dialects (Egyptian vs Gulf = Spanish vs Portuguese) • Right-to-left script (your tokenizer wasn't built for this) • 12.3 million possible word forms vs 170K in English • Diacritical marks change meaning (often omitted) • MSA vs colloquial dialects—models need both **The Cultural Context Problem** Western LLMs are trained on 90%+ English text with Western cultural norms. When analyzing Arabic poetry, they miss: • 1,400 years of Islamic literary tradition • Pre-Islamic poetry conventions • Quranic references and hadith allusions • Arab cultural values (honor, hospitality, family) Example: I asked GPT-4 to analyze a poem about longing. It interpreted "desert journey" literally. Every Arab student knows this is a metaphor for spiritual seeking. **The Business Reality** 420M Arabic speakers. $3T GDP across MENA. Your AI accuracy drops 40% when they use it. This isn't technical—it's market access. Every accuracy point lost is market share given to Arabic-first competitors. **The Path Forward** 1. Arabic-specific tokenizers (not English BPE) 2. Culturally-aware training data (not translated English) 3. Dialect-aware models (not one-size-fits-all MSA) 4. Human-in-the-loop cultural validation The winners won't translate English models. They'll rebuild from the ground up with Arabic linguistic and cultural reality as the foundation. Because "multilingual support" isn't a checkbox. It's a commitment to linguistic and cultural competence most Western AI teams have never built. #ArtificialIntelligence #NLP #ArabicAI #MultilingualAI #EnterpriseTech #LanguageModels #SwalihAI

1