AI-Specialist

Must Have (6+ years, of which 3+ in production LLM systems)

• Production deployment of open-source LLMs — Llama, Falcon, Jais, Mistral, or equivalent — on self-hosted GPU infrastructure
• vLLM, Text Generation Inference, or equivalent high-throughput serving frameworks
• RAG architecture in production: chunking strategies, embedding models, retrieval ranking, reranking, query rewriting
• Vector databases — pgvector, ChromaDB, Weaviate, Qdrant, or Milvus (pgvector and ChromaDB are the target stack)
• Embeddings for bilingual / multilingual content with RTL handling
• Fine-tuning and parameter-efficient fine-tuning (LoRA, QLoRA, PEFT) on transformer models
• LLM evaluation frameworks — RAGAS, TruLens, or equivalent — and ability to design custom evaluation harnesses
• Python, PyTorch, Hugging Face Transformers, LangChain or LlamaIndex
• NVIDIA GPU operations — CUDA awareness, multi-GPU inference, NVLink topology, model parallelism

Strongly Preferred

• MCP (Model Context Protocol) server development experience
• Experience with UAE-origin models — Falcon, Jais, Jais Chat — or close equivalents
• Arabic NLP — tokenisation, normalisation, dialectal handling (Gulf, Saudi, MSA), diacritic management
• Prompt engineering at production scale, including chain-of-thought, few-shot, and instruction templates
• LLM guardrails — Guardrails AI, NeMo Guardrails, or equivalent — for output filtering, PII redaction, jailbreak prevention
• Kubernetes operation of GPU workloads, including NVIDIA k8s operators
• Document AI — Puppeteer or equivalent for PDF rendering with bilingual RTL/LTR typography