Pantheion AI

AI Engineer — The Core (Pantheion-1)

Abu Dhabi, UAEFull-TimeOn-sitePosted Apr 6, 2026

Distributed TrainingRLHFPythonPyTorchHuggingFaceAir-gapped Deployment

Role overview

You will be a core engineer on Pantheion-1 — our frontier Arabic/English large language model, trained on sovereign GCC data and deployed exclusively in air-gapped government cloud environments. This role sits at the technical frontier of sovereign AI: building, evaluating, and deploying the most capable Arabic-native LLM in existence, within the most demanding classified infrastructure constraints. You will work across the full model development stack — from pre-training data pipelines and distributed training to fine-tuning, alignment, and production deployment in air-gapped Gov-Clouds.

What you will do

Design and implement distributed pre-training pipelines for Pantheion-1 on 100–200 MW of classified sovereign compute — optimizing throughput, fault tolerance, and training efficiency across thousands of GPUs
Build and maintain Arabic-language data curation pipelines: sourcing, cleaning, deduplication, quality filtering, and annotation at billion-token scale across Gulf, Levantine, and MSA Arabic sources
Implement RLHF and Constitutional AI alignment pipelines adapted for GCC cultural, religious, and regulatory alignment requirements
Design and execute model evaluation: Arabic NLP benchmarks, safety evaluations, classified-context task performance, and government procurement milestone testing
Engineer the model serving and inference stack for air-gapped Gov-Cloud deployment — optimizing for latency, throughput, and hardware efficiency within classified infrastructure constraints
Implement secure model versioning, update pipelines, and rollback mechanisms compatible with classified environment change control requirements
Collaborate with the red-team operations and interpretability lab to identify and remediate model safety and alignment issues before government deployment
Develop and maintain engineering documentation and technical runbooks that meet classified programme security standards

Skills profile

Required skills

Distributed TrainingRLHFPythonPyTorchHuggingFaceAir-gapped Deployment

Required qualifications

Domain knowledge

5+ years of AI/ML engineering experience with at least 2 years directly working on large language model training or fine-tuning at scale
Strong expertise in distributed training frameworks (Megatron-LM, DeepSpeed, FSDP, or equivalent) and large-scale GPU cluster management
Hands-on experience with RLHF pipelines, instruction tuning, and model alignment methodology
Deep proficiency in Python, PyTorch, and the HuggingFace ecosystem
Experience deploying AI models in air-gapped, offline, or classified compute environments
Ability to obtain UAE security clearance

Preferred qualifications

Bonus domain experience

Direct experience with Arabic NLP — tokenization, morphological processing, dialect-aware training data design
Familiarity with Constitutional AI, RLAIF, or mechanistic interpretability techniques
Experience with model quantization, pruning, and inference optimization for classified edge or on-premises deployments
Prior work in government, defense, or sovereign AI environments
Published research in LLM training, Arabic NLP, or AI alignment