ZTABS fine-tunes large language models for Riyadh enterprises — adapting foundation models to understand Saudi business Arabic, comply with NDMO frameworks, and deliver production-grade performance on the specific tasks your organization needs. From government procurement documents to banking correspondence in Gulf Arabic.
Off-the-shelf LLMs fail in Riyadh for a reason most vendors do not discuss openly: they do not understand Saudi Arabic. GPT-4 and Claude handle Modern Standard Arabic (MSA) reasonably well, but Saudi government procurement documents use a specific bureaucratic register filled with legal terms borrowed from Sharia jurisprudence. Banking correspondence blends Gulf Arabic with English financial terminology. Healthcare records from KFSH&RC or Saudi German Hospital mix Arabic clinical notes with Latin medical nomenclature. Fine-tuning solves this by training a base model on your organization's actual data — the documents, emails, reports, and transcripts that define how your business communicates. ZTABS specializes in Arabic-first LLM fine-tuning: we have benchmarked every major open-weight model on Saudi-specific text and know which architectures respond best to Arabic domain adaptation.
Riyadh's LLM fine-tuning demand is driven by the gap between what commercial AI products offer and what Saudi organizations actually need. SDAIA (Saudi Data & AI Authority) has set aggressive AI adoption targets under Vision 2030, but off-the-shelf solutions consistently underperform on Arabic tasks. Consider the specifics across sectors. Government: Saudi government procurement documents use a specific bureaucratic register filled with legal terms from Sharia jurisprudence — terms like كفالة (kafala/guarantee), استدراج عروض (invitation to bid), and ترسية (award) that do not appear in standard Arabic NLP training data. The Ministry of Justice's case documents follow Hanbali legal traditions with vocabulary absent from any commercial LLM. Royal decrees, Council of Ministers decisions, and Shura Council recommendations use formal Arabic that differs significantly from newspaper Arabic or conversational Gulf dialect. SDAIA itself needs fine-tuned models for the National Data Bank — classifying and extracting entities from millions of government records spanning decades of Saudi administrative history. Energy: Saudi Aramco's technical reports blend Arabic prose with English engineering specifications — a format that confuses zero-shot LLMs. Downstream from Aramco, SABIC's chemical safety data sheets mix Arabic regulatory language with IUPAC chemical nomenclature. Saudi Electricity Company's grid operations reports combine Arabic operational notes with English technical metrics. Ma'aden's mining exploration reports integrate Arabic geological descriptions with international mining classification codes (JORC, NI 43-101). Each of these organizations generates proprietary text data that requires domain-specific fine-tuning for AI to be useful. Banking and finance: SAMA (Saudi Central Bank) regulatory circulars use Arabic financial terminology with no English equivalent — terms like التعثر الائتماني (credit default in Islamic banking context) carry specific Sharia implications that generic translation misses. Al Rajhi Bank processes millions of Arabic customer communications annually — complaint classification, request routing, and sentiment analysis all require models that understand Saudi banking Arabic, including Gulf dialect variations common in customer-facing channels. The Islamic finance dimension is particularly challenging: fatwa documents, Sharia board rulings, and murabaha/ijara contract structures use Arabic legal-financial terminology that exists in no standard training corpus. Healthcare: KFSH&RC's clinical notes mix Arabic clinical descriptions with Latin medical nomenclature and English abbreviations. Saudi German Hospital Group's patient records across 14 facilities use inconsistent Arabic medical terminology — the same condition described differently across hospitals. The Ministry of Health's disease surveillance reports require NLP that understands Arabic epidemiological language. Pharmaceutical marketing materials must comply with SFDA (Saudi Food and Drug Authority) Arabic labeling requirements — fine-tuned models for regulatory compliance checking. Education: Saudi universities processing thousands of Arabic academic papers, dissertations, and research proposals need fine-tuned models for Arabic academic text classification, plagiarism detection in Arabic, and Arabic-to-English research translation. NEOM Academy and the new universities in NEOM require Arabic-English bilingual AI for curriculum development and student assessment. Jais, the Arabic LLM developed by G42 in Abu Dhabi, was a breakthrough — but Jais is a general-purpose model, not fine-tuned for your specific domain. The real value comes from taking Jais (or Llama 3, or Command R+) and training it on your organization's proprietary corpus. The result is a model that speaks your company's language, not generic Arabic.
Our Riyadh fine-tuning engagements start with data audit, not model selection. We catalog the client's text data: volume, language distribution (Arabic vs. English vs. mixed), domain vocabulary, and quality issues (OCR errors from scanned documents are endemic in Saudi government workflows). We classify data sensitivity under NDMO's Data Management Standards — most fine-tuning datasets contain personal or confidential information that cannot leave KSA. Training infrastructure runs on-premise or on KSA-region cloud (AWS Bahrain, Azure UAE). For Arabic-specific fine-tuning, we have found that LoRA adapters on Jais 30B or Command R+ deliver the best Arabic performance for classification and extraction tasks, while Llama 3 70B with full fine-tuning produces superior Arabic text generation. We use Weights & Biases for experiment tracking (self-hosted instance to keep metrics within KSA) and run systematic evaluations against human-annotated Saudi Arabic test sets — not translated English benchmarks, which consistently overstate model quality on Arabic. Every engagement includes a red-team phase where native Arabic speakers stress-test the model with dialectal variations, code-switching, and edge cases specific to the client's domain.
Government Procurement
A Saudi government entity processing: 3,000+ procurement contracts annually (total value: SAR 12B+) needed to: automate contract review — extracting key terms, identifying non-standard clauses, and: flagging compliance issues against: GTPL (Government Tenders and Procurement Law) requirements. The problems: (1) Language complexity: contracts mixing: formal Arabic legal language (MSA with: Sharia-derived terminology — كفالة (kafala/guarantee), ضمان حسن التنفيذ (performance bond), غرامة تأخير (delay penalty)), English technical specifications (engineering and: IT procurement appendices), and: Saudi-specific regulatory references (GTPL articles, Etimad platform codes). No: off-the-shelf NLP tool: handled this: trilingual-domain complexity. (2) Volume: 3,000+ contracts annually requiring: extraction of: 47 distinct clause types, identification of: commercial terms (value, duration, penalties, performance bonds, payment schedules), and: comparison against: GTPL standard templates to: flag deviations. Manual review: consuming 12,000+ staff hours per year across: 15 legal reviewers. (3) Sovereignty: procurement data classified: Level 3 under NDMO framework — cannot leave: government infrastructure under: any circumstances.
We fine-tuned a Jais 30B model on: 8,000 historical procurement contracts annotated by: the entity's senior legal team over: 3 months of: labeling workshops. The model: learned to: identify 47 distinct clause types, extract key commercial terms (value, duration, penalties, performance bonds, insurance requirements, subcontracting limitations), flag deviations from: GTPL standard templates (highlighting: the specific clause and: the corresponding GTPL article), and: generate Arabic-language: summary reports for: each contract. A separate Llama 3 8B model: handles the: English technical appendices — extracting: technical specifications, warranty terms, and: SLA commitments. Both models: run on: the entity's private infrastructure — 4× NVIDIA A100 GPUs in: their Riyadh data center with: no internet connectivity.
Contract review time: from 4.5 hours to: 35 minutes per contract (87% reduction). Non-standard clause detection: 91.3% precision (validated against: 500 contracts reviewed by: senior legal staff). The model: catching: 23 potentially problematic clauses in: the first quarter that: manual reviewers had: previously missed — clauses that: could have: exposed the entity to: financial risk. Data sovereignty: zero procurement data: leaving the entity's infrastructure since: deployment. NCA security audit: passed without: findings. Staff impact: 9,500 staff hours saved: in the first year — legal reviewers: redeployed from: manual extraction to: strategic contract negotiation and: vendor relationship management. The entity: now processing: 4,200 contracts annually (40% increase in: throughput) with: the same headcount — the AI: removing: the review bottleneck that: previously constrained: procurement velocity.
We clean, deduplicate, and structure your training data into high-quality instruction-response pairs. Quality data is the single biggest factor in fine-tuning success.
Fine-tune GPT-4o Mini and GPT-3.5 Turbo through OpenAI's API with systematic hyperparameter optimization, validation splits, and automated evaluation.
Fine-tune Llama 3, Mistral, Phi, and other open-source models using LoRA, QLoRA, and full fine-tuning on cloud GPUs or your own infrastructure.
Rigorous evaluation against your specific tasks with automated benchmarks, human evaluation, and A/B testing against base models to quantify improvement.
Align model outputs with human preferences using DPO (Direct Preference Optimization) and RLHF techniques for better quality and safety.
Deploy fine-tuned models via OpenAI, vLLM, TGI, or Ollama with optimized inference, batching, and auto-scaling for production workloads.
Riyadh's LLM fine-tuning market is nascent but growing fast. (1) G42 ecosystem (Abu Dhabi) — G42 developed Jais and offers enterprise AI through Inception, but focuses on large sovereign contracts (SAR 10M+ minimum engagements). Presight (G42 subsidiary) provides analytics but not custom fine-tuning services. For mid-market Saudi enterprises, G42 is oversized and overpriced. (2) Saudi AI companies — SCAI (Saudi Company for AI) is building local capability but remains early-stage. Mozn (Riyadh, backed by STV) builds Arabic NLP products (Nar platform) but operates as a product company, not a fine-tuning services provider. Lucidya (Riyadh) focuses on Arabic social media analytics. These companies serve specific niches, not general enterprise fine-tuning. (3) Global consultancies — McKinsey, BCG, Bain, and PwC all have KSA AI practices. They produce AI strategies and model selection recommendations (SAR 2-5M engagements) but subcontract the actual ML engineering to offshore teams or local integrators. The fine-tuning work itself is rarely done in-house. (4) Academic partnerships — KAUST produces world-class NLP research (including Arabic NLP papers at ACL and EMNLP), but academic projects prioritize research novelty over production reliability. KFUPM and KSU have growing AI programmes but limited enterprise deployment experience. (5) Offshore development — Indian firms offer cost-competitive ML development (SAR 400-700/hr) but consistently underdeliver on Arabic NLP quality. Fine-tuning Arabic models requires native linguistic understanding — Arabic morphology, dialectal variations, and domain-specific terminology cannot be handled by engineers who learned Arabic from training data. The quality gap is most visible in evaluation: offshore teams report high accuracy on standard Arabic benchmarks but underperform on Saudi-specific test sets. ZTABS fills a specific gap: production-grade Arabic LLM fine-tuning with native Arabic linguistic expertise, the infrastructure to run the full pipeline within Saudi borders, and the MLOps capability to maintain fine-tuned models in production.
LLM fine-tuning projects in Riyadh range by scope and complexity. Single-task fine-tuning (classification, extraction, or summarization for one domain, 4-6 weeks): SAR 150,000-250,000 ($40,000-$67,000) — the most common starting point, covering data preparation, model selection, fine-tuning, evaluation, and production deployment. Multi-task domain platform (multiple Arabic NLP tasks on shared model infrastructure, 8-12 weeks): SAR 350,000-600,000 ($93,000-$160,000) — for organizations needing classification, extraction, generation, and question-answering across their Arabic document corpus. Enterprise Arabic AI platform (custom models, RAG integration, multi-department deployment, 12-18 weeks): SAR 700,000-1,200,000 ($187,000-$320,000) — full-scale deployment across the organization. The government procurement case study (Jais 30B fine-tuning with 8,000 annotated contracts) fell in the multi-task range. Data annotation — typically 30-40% of project cost — varies significantly based on document complexity and required expertise level (legal annotation by qualified lawyers costs 3-5x more than general text annotation). GPU infrastructure for fine-tuning runs SAR 8,000-35,000/month during the training phase. Senior ML engineers with Arabic NLP experience command SAR 55,000-85,000/month locally — among the scarcest tech professionals in the Kingdom. ZTABS delivers at SAR 130-180/hr with deep Arabic fine-tuning expertise and KSA data residency capability built in.
Verified reviews from real client engagements — sourced from our public testimonial archive and Clutch profile.
My experience is throughout positive. Communication, service, the short response times and the flawless execution of a challenging topic was absolutely great. ZTABS is definitely my first choice again.
Christian Neff
Bank Software Advisory · Bank Software Advisory
Fintech
Fantastic Agency! I couldn't fault them even if I tried. They always go above and beyond to meet your expectations and always produces quality work. Thank you ZTABS.
Stephanie Kal
CEO · Beauty Finder Australia
Marketplace
It has been great working with ZTABS. They bounce off the ideas along the way. Amazing Experience.
Joel Rowe
CEO · Drill Quoter
Marketplace
We don't just contract. We ship and operate our own software. 23 products in production.
Common questions about LLM fine-tuning for Riyadh businesses
It depends on the task. For text classification and entity extraction in Arabic, Jais 30B with LoRA adapters delivers the best accuracy per compute dollar. For Arabic text generation (reports, summaries, correspondence), Llama 3 70B with full fine-tuning outperforms Jais on fluency. For bilingual Arabic-English tasks, Command R+ handles code-switching well. We benchmark all candidates against your actual data during a 2-week evaluation phase before committing to an architecture.
ZTABS is a remote-first LLM fine-tuning agency serving Houston businesses — including data pipeline & curation, openai fine-tuning, open-source model training. We work with Energy & Oil/Gas, Healthcare & Biotech, Aerospace & Defense companies in Houston, TX via timezone-aligned engineers and async workflows; we do not have a local office, and we are explicit about that with every client.
LLM Fine-Tuning in New York, NYZTABS is a remote-first LLM fine-tuning agency serving New York businesses — including data pipeline & curation, openai fine-tuning, open-source model training. We work with Finance & Fintech, Media & Advertising, Fashion & Retail companies in New York, NY via timezone-aligned engineers and async workflows; we do not have a local office, and we are explicit about that with every client.
LLM Fine-Tuning in Los Angeles, CAZTABS is a remote-first LLM fine-tuning agency serving Los Angeles businesses — including data pipeline & curation, openai fine-tuning, open-source model training. We work with Entertainment & Media, E-commerce & DTC Brands, Gaming & AR/VR companies in Los Angeles, CA via timezone-aligned engineers and async workflows; we do not have a local office, and we are explicit about that with every client.
Web Development in Riyadh, Saudi ArabiaZTABS is a remote-first web development agency serving Riyadh businesses — including full-stack development, progressive web apps, api development. We work with GovTech, FinTech, Energy Tech companies in Riyadh, Saudi Arabia via timezone-aligned engineers and async workflows; we do not have a local office, and we are explicit about that with every client.
Web Design in Riyadh, Saudi ArabiaZTABS is a remote-first web design agency serving Riyadh businesses — including ui/ux design, responsive design, custom interfaces. We work with GovTech, FinTech, Energy Tech companies in Riyadh, Saudi Arabia via timezone-aligned engineers and async workflows; we do not have a local office, and we are explicit about that with every client.
AI Development in Riyadh, Saudi ArabiaZTABS is a remote-first AI development agency serving Riyadh businesses — including llm integration & fine-tuning, ai agents & automation, rag & knowledge systems. We work with GovTech, FinTech, Energy Tech companies in Riyadh, Saudi Arabia via timezone-aligned engineers and async workflows; we do not have a local office, and we are explicit about that with every client.
LLM Fine-Tuning ServicesLearn more about our LLM fine-tuning services nationwide.
Partner with ZTABS for expert LLM fine-tuning in Riyadh. Get a free consultation today.
