Gen AI Engineer (LLM &RAG

Urgent requirement for Gen AI Engineer (LLM &RAG) in Banking Domain

is required for our banking clients in Kuwait

• Strong in Azure OpenAI Service, Azure AI Search, Azure AI Content Safety, Azure AI Foundry (prompt flow/evaluation), Copilot Studio..
  
  -
• Hands on Python with Lang Chain, Llama Index, Semantic Kernel;
  Hugging Face Transformers; open embedding models..
  
  --
• Hands on pgvector, Qdrant, Chroma, Milvus; evaluation tools like RAGAS, Deep Eval.
  
  -
• Guardrails (Azure AI Content Safety, Llama Guard, NeMo Guardrails), retrieval‑time access control, PII/PCI masking, audit trails.

The builder of the intelligence behind every GenAI and Agentic AI use case . This role designs and implements the retrieval-augmented generation pipelines — chunking, embeddings, retrieval, re-ranking, prompting and grounding making sure AI return accurate, sourced, role‑aware answers. Owns prompt design, guardrails, evaluation harnesses and (where needed) fine‑tuning or domain‑tuning of models. Ensures answers are traceable to source documents, a hard requirement for compliance and trust in a bank.

4+ years in ML/NLP or software engineering, with 1.5+ years hands‑on building LLM / RAG applications. Proven delivery of a RAG system: document ingestion, embeddings, vector search, prompt orchestration and evaluation. Experience with hallucination control, grounding, citation of sources, and structured evaluation of GenAI quality. Familiarity with fine‑tuning / domain adaptation and with prompt‑injection and jailbreak defense.

Microsoft stack (reference build):
Azure OpenAI Service, Azure AI Search, Azure AI Content Safety, prompt flow and evaluation in Azure AI Foundry. Copilot Studio for conversational experiences on the KIB intranet.

Python with Lang Chain / Llama Index / Semantic Kernel;
Hugging Face Transformers and open embedding models. Open vector databases (pgvector, Qdrant, Chroma, Milvus); open evaluation tooling (RAGAS, Deep Eval); local serving with Ollama / vLLM. Open and fine‑tunable models (Llama, Mistral) with LoRA / PEFT techniques.

Implements guardrails and content filtering (Azure AI Content Safety or open equivalents such as Llama Guard / NeMo Guardrails) against prompt injection, data leakage and unsafe output. Enforces retrieval‑time access control so a user only ever sees content their identity is entitled to (document‑level and row‑level security passed from Entra  / the source system). Prevents sensitive data (PII, client, PCI) from being logged or sent to models outside the approved boundary;

applies masking and redaction in the pipeline. Builds source‑citation and audit trails so every answer can be traced to approved material — essential for regulatory defensibility.

Skills:

llm,rag,genai