Goal
Public and private credit investor manager was looking to streamline its credit risk process with an interactive investment knowledge chatbot to assist analysts to extract and interrogate insights
Approach
We mapped the investment process to define requirements for decision thresholds, UI, workflow, response time and data management.
We build a conversational chat capability with a vector database for a given set of files, supplemented by a chat history to support follow-up questions.
This was extended with a PDF API to load temporary files for Q&A to support ad-hoc or independent topic research.
We developed a proprietary task agent with the ability to generate, execute and adapt a transparent set of actions to answer complex questions.
This was enhanced with a Chain-of-Thought based agent with the ability to search and dig deeper into pages to retrieve information available online.
Our co-pilot was deployed with the UI, API, and DB containerised and distributed in mono-repo structure.
We utilised a Redis database with in-memory data store for vector and text-based searches using HashMaps plus a FastAPI cache to speed up application with fast response from static API endpoints like /info paths.
Outcome
We delivered a chat-based co-pilot workflow with the following capabilities:
- Summarization and key point extraction from multi-modal data sources and documentation
- Conversational interface ability to rephrase a question and create a plan to improve accuracy
- Planning agent with the ability to create a plan leveraging multiple tools, steps and sources -- and self-correct when needed
In addition, we up-skilled the internal data science team on technical best practices through day-to-day collaboration and on-the-job coaching.