Sector
Insurance and pensions
Continuous testing and quality assurance for live models in insurance
The customer quotes and brand details below are representative content, prepared to show the structure of the template.
The challenge
There was no way to check by hand whether the live policy assistant's answers were accurate, and no way to measure the side effects a model update introduced.
Our approach and integration
On the principle that an unmeasured system does not go live, we built a CI/CD evaluation pipeline on Ragas and DeepEval. Wired into the company's Jenkins and GitHub, it scores faithfulness, answer relevance and latency automatically on every code or prompt change.
Results and impact
The rate of wrong answers and hallucinations in production fell from 12% to 0.2%.
Cutting needless, over-long prompts saved 40% of token spend.
QA and test cycles came down from three weeks to two hours.
NEXT
What Could We Achieve With Your Team?
See in practice, not in theory, the momentum our solutions can give your organisation.