MUFG Bank

Solution

MUFG’s AI/ML team built a production RAG application that digests corporate data and turns it into sales-ready presentation material. In the research phase, the team used the Python version of LangChain with Streamlit to stand up chat and retrieval quickly, while the Retriever interface allowed different vector databases and search engines to be swapped and compared at low implementation cost. The system extracts key facts from long annual reports, market data, and disclosures, then uses fine-tuned prompting, retrieval, and few-shot guidance to surface the most relevant sections and generate structured recommendations. In production, MUFG moved to the TypeScript version of LangChain on Next.js, where Runnable Lambda dynamically changes content filters and indexes inside a custom RAG chain. That architecture turns research-heavy sales prep into a reusable knowledge workflow rather than a specialist-only manual process.

Data flow

Annual reports, market data, and disclosures are ingested into retrieval back ends chosen through the Retriever interface, LangChain extracts and summarizes the most relevant sections, few-shot prompting turns those insights into presentations and recommendations, and the production Next.js app serves the outputs to corporate sales teams.