LLM

Over the past few years, LLMs (Large Language Models) have had a transformative impact on how software is built and how businesses automate complex work, to the point where they merit dedicated consideration rather than being treated as just another tool.

I’ve used LLMs extensively as part of agentic AI systems, including LangGraph-powered workflows that support dynamic decision making, branching logic, and long-running processes. In these systems, LLMs are not used in isolation, but as reasoning components within structured graphs that combine deterministic code, tool use and memory. This approach allows AI systems to adapt to changing inputs, recover from errors, and make context-aware decisions while remaining observable and controllable. Evaluation matters as much as the build. I’ve benchmarked LLM techniques against golden datasets and used models as judges for unstructured outputs, since a system that can’t be measured can’t be trusted in production.

LLMs have already demonstrated significant real-world value across analysis, automation, and user interaction, and their trajectory is clear. Model capability and reliability continue to improve, while costs fall as research advances in efficiency, inference optimisation, and architecture design. When paired with the right engineering patterns and guardrails, LLMs are becoming a practical foundation for production systems rather than experimental technology.