Artificial Intelligence Engineer

Xpin AI; Dubai; On-site; AED 420,000 - 630,000; Full time

Job summary

Xpin AI, based in Dubai, seeks a senior engineer to own the agentic layer of products, building the full stack from async streaming backend to real-time UI. You will work closely with clients, participate in on-premise installations, and drive production-ready, multilingual AI initiatives across fixed hardware and cloud environments.

You will lead architecture for multi-agent orchestration, tool calls, and retrieval-grounded responses, with a focus on reliability, observability, and scalable

Qualifications

• shipping agentic AI systems to production with latency, throughput and cost considerations
• design multi-agent, tool-use, planning and recovery mechanisms
• real-time UI and streaming backend for AI systems
• on-premise and cloud deployments including air-gapped environments
• experience with self-hosted open-weight models and serving stacks

Responsibilities

• Own the agentic layer of products and build full stack from async backend to real-time UI
• Lead on-premise installation, integration, and client conversations
• Ensure production reliability across on-premise and cloud deployments
• Maintain evaluation practices to ground truth and monitor hallucination/quality
• Tune retrieval layers (Graph-RAG, local embeddings) and latency budgets

Skills

LLMOps

Agentic systems design

Full-stack AI systems

On-premise deployment

Tools

LangGraph

LangChain

Temporal

vLLM / LLM serving

Job description

Design and ship production agentic AI systems that reason over real enterprise and government data, then build the streaming, real-time interfaces that make them usable, deployed across both air-gapped on-premise environments and cloud.

About Xpin AI

Xpin AI builds sovereign AI for governments, enterprises, and institutions across the GCC. The work is multilingual at its core, spanning Arabic, English, and regional languages, and it runs across on-premise sovereign environments and cloud. What makes the engineering interesting is the constraint, doing frontier-grade agentic and multilingual AI with no data egress, on fixed hardware, for customers who expect it to work the first time. If you like hard problems with real users on the other side, you will not be bored here.

The role

This role is based in our Dubai office. You will own the agentic layer of the products you are assigned to and build the full stack around it, from the async streaming backend to the real-time UI. Because you are close to the client environments we deploy into, you will also be involved in on-premise installation, integration, and the technical conversations that come with them, which is the part of the job you cannot do from behind a laptop alone.

You will work in a squad in an Agile setup, reporting to a Lead AI Engineer or Solutions Architect, with genuine ownership of the systems you build rather than tickets handed down. By the end of your first year, success looks like agentic systems you designed running reliably in production across both on-premise and cloud deployments, grounded by a retrieval layer you tuned, and kept honest by an evaluation practice you helped stand up.

What you'll build

• Design and ship production agentic systems: multi-agent orchestration, tool and function calling, planning loops, and multi-hop reasoning, with the state management, recovery, and guardrails that keep them reliable in production rather than impressive in a demo.
• Build the Model Context Protocol servers and clients that let agents act safely against internal systems, on-premise against air-gapped data and on cloud for connected projects.
• Build the advanced retrieval that grounds the agents: hybrid and graph-based search, query routing, local embeddings, and reranking, tuned so answers are accurate rather than merely plausible.
• Stand up the evaluation and observability practice from scratch: offline trajectory evals, tracing, hallucination and response-quality monitoring, and regression gates that turn "it seems to work" into something measured.
• Fine-tune open-weight multilingual models on customer data and own the serving path, including quantization and VRAM budgeting on local H200 servers.
• Build the AI-native interfaces that make these systems usable: low-latency token-by-token streaming, real-time voice interactions, and context-aware components, served over WebSockets and Server-Sent Events.
• Support on-premise deployment and integration in client environments, including offline packaging, GPU sizing, and post-deployment debugging.
• Own latency, throughput, and cost across both deployment targets, where the on-premise hardware is fixed and the cloud bill is not.

Tech stack

• Models and serving: self-hosted open-weight models (Llama, Qwen, Mistral and Mixtral, DeepSeek) served with vLLM, TGI, TensorRT-LLM, or Ollama; quantization (AWQ, GGUF, FP8); hosted models (Azure OpenAI, Claude, Gemini) on confirmed connected projects.
• Agents and orchestration: LangGraph, LlamaIndex, LangChain, CrewAI, AutoGen; durable execution with Temporal; Model Context Protocol (MCP); planning patterns such as ReAct and plan-and-execute.
• Retrieval: pgvector or Qdrant; hybrid and Graph-RAG; query routing; local embeddings (bge, e5); reranking; programmatic prompt optimization (DSPy).
• Fine-tuning: PyTorch, LoRA and QLoRA, PEFT, Axolotl or Unsloth.
• Evaluation and observability: Promptfoo, Langfuse, Phoenix, LangSmith, TruLens.
• Multi-modal and voice: streaming STT and TTS, self-hosted (faster-whisper, Piper) on air-gapped projects and hosted (Azure Speech, Deepgram, ElevenLabs) on connected ones.
• Backend: Python with async FastAPI and Pydantic; streaming APIs over WebSockets and Server-Sent Events.
• Frontend: React 19+, Next.js (App Router), TypeScript, Tailwind; state with Zustand or Jotai; concurrent rendering and Suspense; Web Workers; Bun.
• Data and infrastructure: PostgreSQL and CosmosDB; Docker and Kubernetes (RKE2 or k3s on-premise, AKS on Azure); Harbor and MinIO; CI/CD with GitHub Actions; infrastructure as code with Terraform or Azure Bicep; on-premise GPU servers (NVIDIA H200); Azure as the primary cloud target (App Service, AI Search, Key Vault).

Level and scope

Mid-to-senior, with roughly five or more years building production software and applied AI as a guide rather than a gate. Given the breadth here, agentic systems plus genuine full-stack plus the deployment layer, the engineers who fit will tend toward the senior end, and that is fine. This is an individual-contributor role with real ownership, not people management. For newer tools such as MCP, agent frameworks, and current serving stacks, demonstrated projects and strong fundamentals count more than a year count that does not yet exist.

Must-have

• Has shipped LLM-powered or agentic systems to production and owns the latency, throughput, and cost tradeoffs that come with them.
• Can design multi-step, multi-agent systems with tool use, planning, state, and recovery, not just single-turn prompting.
• Genuine full-stack ability: builds and ships both the async streaming backend and the real-time UI around an AI system, without waiting on a separate hire for either half.
• Has fine-tuned and deployed open models within secure boundaries and understands the serving path that follows.
• Comfortable operating where there is no internet: offline dependencies, self-hosted everything, and GPU sizing against fixed hardware.
• Treats evaluation as part of building, not a thing done later.
• Based in the UAE or willing to relocate to Dubai, and comfortable working on-site.
• Advanced RAG at depth: Graph-RAG, hybrid search, query routing, DSPy, and vector search at scale.
• Real-time voice and multi-modal pipelines, including low-latency conversational interfaces.
• LLMOps and observability practice using tools like LangSmith, Phoenix, or TruLens to monitor hallucination and quality in production.

Nice-to-have

• Azure depth (Azure OpenAI, AI Search, CosmosDB, AKS, Bicep) and infrastructure as code.
• Multilingual NLP, especially Arabic or other regional languages. Arabic language ability is a plus but not required.
• Prior delivery in the GCC, or in air-gapped or otherwise regulated environments.
• Behavioral or predictive intelligence work: sentiment, intent, churn, or forecasting over large unstructured datasets.
• Domain experience in GovTech, BFSI or logistics, product or startup experience, or open-source contributions in the agent and inference ecosystem.

Why join

You will work on genuinely hard problems, air-gapped deployment and on-premise inference of agentic and multilingual AI, on current frontier hardware, with a modern self-hosted stack rather than a thin wrapper around someone else's API. The work ships into real government and enterprise environments across the GCC, so what you build gets used, not shelved. You will sit close to the decisions, with mentorship from senior AI engineers and solutions architects, a learning budget, and room to go deep on problems most teams never touch.

Compensation and logistics

On-site in Dubai. Compensation is competitive and benchmarked to your experience and the UAE market, paid tax-free. Package includes UAE employment visa sponsorship, medical insurance, and annual leave and flight allowance per UAE labour law and company policy. Relocation support is available for candidates joining from outside the UAE.