Overview
AI Engineer - Defence RAG Systems ( Security Clearance Essential ) On Site 2 X Days a week, Plymouth
Clearance: Active SC Essential | Sector: Defence
Role Overview
Defence client requires an SC Cleared AI Engineer to build fully on-premises RAG systems using open-source technologies. You'll develop classified AI capabilities on air-gapped infrastructure with zero external dependencies.
Responsibilities
Build end-to-end RAG pipelines on isolated defence networks using open-source LLMs (Llama 3, Mistral, Qwen)
Deploy local vector stores (Chroma, FAISS, Milvus) with sensitive document ingestion pipelines
Host and optimise LLMs using vLLM/TGI on local GPU clusters without internet connectivity
Implement agent orchestration using LangChain/LangGraph in completely offline environments
Design secure document processing for classified materials with appropriate data sanitisation
Build monitoring and evaluation systems that operate within air-gapped infrastructure
Essential Requirements
Active SC Clearance (non-negotiable) - willingness to undergo DV if required
Demonstrable experience deploying open-source LLMs (Llama, Mistral, Falcon) on-premises
Expertise with local vector databases (Chroma, FAISS, Weaviate) in offline deployments
Strong vLLM/Text Generation Inference experience for high-throughput model serving
Proven ability to work on air-gapped systems with no external package repositories
Experience with GPU orchestration (NVIDIA A100/H100) and CUDA optimisation
Python expertise with offline dependency management and local package mirrors
Technical Stack (All On-Premises)
Models: Llama 3, Mistral, Qwen (locally hosted)
Vector Stores: Chroma, FAISS, Milvus
Orchestration: LangChain, LangGraph for agents
Hosting: vLLM, TGI, Ollama on bare metal/private cloud
Infrastructure: Air-gapped Kubernetes, local container registries
Desirable Skills
Experience with defence/government IT security protocols
Knowledge of CIS benchmarks and NCSC guidelines
Familiarity with cross-domain solutions and data diodes
Understanding of classification marking and handling procedures
#J-18808-Ljbffr