The domain in detail

LLM systems in production

For organisations that put language models into a running process. The same approach as Esmond, which runs in four countries.

Schedule a call
What we build
  • Language model pipelines in several stages, each with its own instruction and failure path
  • Assessment of files, applications and documents against a fixed standard
  • Deviations flagged with their location in the source
  • No grade, no decision
How it is built
  • Anthropic, OpenAI, AWS Bedrock, Google Vertex AI
  • Prompt architecture and prompt caching
  • Fallback between models: the failure of one part does not stop the run
  • Findings that cannot be traced back to the text are dropped
  • Cost per run measured and capped
  • Model changes through A/B, with a quality check before rollout
How it is developed
  • Claude Code and Codex in daily work
  • Own agents for review, compliance and cost
  • Tests before every merge, smoke test after every deploy
Suited and not suited
  • Suited: streams with many files and limited stakes per file
  • Not suited: the single file with high stakes
  • Never as a replacement for human judgement
Contact
Back to the homepage