ClearBank

Solution

ClearBank describes an in-house generative AI implementation for regulated banking API workflows built around Large Language Models (LLMs) and retrieval augmented generation (RAG) architecture. The system ingests internal policies, policy and client documents, and client transactional enquiries or complaints into separate pipelines for enquiry monitoring, policy review, and client screening. RAG is used to ground outputs in source documents, while the generative models run on ClearBank's internal infrastructure rather than an external shared service. An evaluation pipeline continuously measures document specificity, precision, QA correctness, hallucinations, toxicity, and, for summarisation use cases, ROUGE. A model experiment tracking system records the exact base model version used for each output so end users can trace responses. Domain experts remain in the loop to provide real-time feedback, identify bias, and help product owners tune models and refine input documents under a single assurance approach.

Data flow

Client transactional requests, complaints, internal policies, and policy or client documents are ingested into ClearBank's generative AI pipelines. RAG retrieves supporting source material for the LLMs, which generate monitoring, review, screening, or summarised outputs on internal infrastructure. The evaluation pipeline and model experiment tracking system then measure quality, log model versions, and support end-user traceability, while domain experts feed back tuning signals.