Artificial Intelligence Developer

Role & responsibilities:

• Design and implement multi-agent orchestration pipelines coordinating sequences of specialised AI agents across a structured, multi-stage workflow
• Implement Human-in-the-Loop (HITL) workflow engine stateful approval gates where named human reviewers must approve, edit, or reject AI outputs before pipeline advancement
• Ensure all agents are stateless between sessions no cross-session memory accumulation, no passive learning from runtime data
• Build and maintain the Model Abstraction Layer decoupling agent logic from specific model versions to enable zero-downtime model upgrades
• Engineer sophisticated multi-turn system prompts encoding domain knowledge, evaluation criteria, output structure requirements, and adversarial review logic
• Implement RAG (Retrieval Augmented Generation) pipelines using embedding models for semantic document retrieval context window management, chunking strategy, and retrieval precision optimisation
• Manage model inference parameters (temperature, top-P, top-K, max tokens) per agent role balancing output consistency and quality
• Integrate with Azure OpenAI (GPT-5.5, GPT-5.4, text-embedding-3-large) and/or Vertex AI (Gemini 2.0 Pro, Gemini 2.0 Flash, text-embedding-gecko) via managed API endpoints
• Build and maintain the deterministic tool layer conventional software components (database queries, REST API calls, file parsers, data validators, schema normalisers) that handle all evidence retrieval and data access with exact, auditable, reproducible results
• Implement read-only service account integrations with enterprise source systems
• Design and implement evidence validation logic completeness thresholds, exception handling, missing evidence escalation procedures
• Implement offline RSA-4096 JWT-based license enforcement cryptographic license validation without network dependency
• Build immutable audit trail logging every agent action, HITL decision, evidence retrieval, and output version logged in structured, tamper-evident form
• Containerise all components for Kubernetes deployment (AKS / GKE) Helm chart authoring, Blue-Green deployment support

Preferred candidate profile:

• 3+ years building production LLM-integrated applications not prototypes or notebooks, but shipped systems handling real data
• Deep expertise in agentic AI patterns ReAct loops, tool use, multi-agent orchestration, agent memory and state management
• Hands-on prompt engineering at production scale system prompt design, few-shot construction, chain-of-thought, structured output enforcement, adversarial/critic prompt patterns
• RAG implementation experience embedding models, vector databases (pgvector, Pinecone, Weaviate, or equivalent), semantic chunking, retrieval evaluation
• Proficiency with Azure OpenAI SDK and/or Vertex AI SDK model invocation, streaming, error handling, retry logic, token management
• Understanding of LLM failure modes — hallucination, confidence calibration, context window limits, prompt injection — and architectural mitigations for each
• Expert-level Python — async programming, Pydantic data models, structured output parsing, robust error handling
• REST API design and integration — building and consuming enterprise APIs (OpenAPI spec, authentication patterns, rate limiting, retry logic)
• Relational database development — PostgreSQL / Cloud SQL schema design, query optimisation, transaction handling
• Cloud-native development on Azure or GCP — containerisation (Docker), Kubernetes (AKS/GKE), managed services (object storage, key vault/secret manager, monitoring)
• Git-based development workflow — trunk-based development, code review, CI/CD pipeline integration
• Experience building within security-constrained environments — understanding of encryption at rest/in transit, secrets management, principle of least privilege, audit logging requirements
• Familiarity with confidential computing concepts (SGX enclaves, AMD SEV, or equivalent) — not required to be a specialist, but must understand the deployment constraints
• Understanding of data residency requirements and BYOC (Bring Your Own Cloud) deployment models