Agentic AI Developer London

As Associate Director of Applied AI/ML at Fitch Ratings, I design and build production agentic AI systems daily. My work involves architecting multi-agent workflows, autonomous research agents, and document intelligence pipelines that enable analysts to interact with credit ratings data, financial filings, and regulatory documents through natural language. This is agentic AI in production, in a regulated environment, at scale.

What Agentic AI Actually Means in Production

Agentic AI is not a chatbot. It is a system where LLMs reason, plan, use tools, and execute multi-step tasks autonomously. In practice at Fitch Ratings, this means:

• Agents that retrieve, synthesize, and summarize information across large financial document corpora
• Multi-agent workflows where specialist agents collaborate to complete research tasks
• Autonomous pipelines that orchestrate tool calls, vector search, and LLM reasoning
• Human-in-the-loop checkpoints for decisions requiring analyst validation
• Robust error handling, retry logic, and observability built for production reliability

Frameworks and Stack

LangChain and LangGraph

LangGraph is the primary framework for multi-agent orchestration at Fitch. Its stateful graph model enables complex agent workflows with conditional branching, parallel execution, and persistent memory across steps.

• ReAct agents for reasoning and tool use
• Plan-and-execute architectures for multi-step research tasks
• Supervisor agents routing work to specialist subagents
• LangChain for document loaders, prompt templates, and LLM integrations

LlamaIndex

Used for advanced RAG patterns, query routing, and knowledge graph construction over financial document corpora.

Foundation Models

• AWS Bedrock for multi-model access and enterprise compliance
• Azure AI Foundry for Azure-hosted model serving
• OpenAI API for GPT model integration

Infrastructure

• FastAPI for agent-serving APIs
• Redis for agent state caching and tool result memoization
• Celery for async agent task execution
• Docker and Kubernetes for containerized deployment
• Airflow for pipeline orchestration

Document Intelligence and RAG

At Fitch, agents work over large-scale document corpora including credit rating reports, financial filings, and regulatory documents. The document intelligence pipeline includes:

• Document extraction with pymupdf4llm for high-fidelity PDF parsing
• Chunking strategies optimized for financial document structure
• Embedding models for semantic representation
• Azure AI Search as the vector store for scalable retrieval
• Hybrid search combining semantic and keyword matching
• Re-ranking for retrieval quality

MLOps and LLMOps for Agents

Agentic systems require rigorous evaluation before production deployment. This includes:

• Experiment tracking for agent configurations and prompt variants
• A/B evaluation frameworks comparing agent versions
• Systematic quality benchmarks across representative query sets
• Production monitoring for latency, cost, and answer quality
• CI/CD pipelines for agent deployment with GitHub Actions

Industries and Use Cases

Financial Services

Deep experience building agentic AI for credit ratings, financial analytics, and regulatory compliance. Understanding of the reliability, auditability, and compliance requirements that financial services AI demands.

Document-Heavy Domains

Any domain with large volumes of unstructured documents is a natural fit for agentic RAG systems. Financial filings, legal contracts, research reports, technical documentation.

Mentoring and Consulting

Open to mentoring engineers and teams building their first agentic AI systems. Whether it is framework selection, architecture review, evaluation strategy, or production readiness, happy to help. Reach out at [email protected].