Reinsurance Automation Platform
- Customer
- UzbekInvest
- Project manager on the customer side
-
- IT Provider
- GlowByte
- Year of project completion
- 2024
- Project timeline
- September, 2023 - August, 2024
- Project scope
- 30 automated workstations
- Goals
- The goal of the project is to provide a system which can cover end-to-end reinsurance underwriting processes and provide automation for routine tasks.
- Project Results
-
We developed, tested, and deployed a system to production that can automate 70% of routine work, resulting in an 80% decrease in the time required to process each reinsurance application. We incorporated a risk assessment feature using large language models (LLMs) to identify risky attributes within the applications. Our end-to-end system receives raw email messages, produces parsed applications with risk assessments, and provides interfaces to complete the full cycle of underwriting.
The uniqueness of the project
Our automation solution utilizes Large Language Models (LLMs) to transform the processing of reinsurance applications by automatically analyzing submissions and evaluating associated reports to identify potential risks in the insured objects. This innovative approach goes beyond standard automation techniques by leveraging advanced natural language understanding to detect nuanced risk factors that traditional methods might overlook. The system significantly reduces processing time, achieving an 80% improvement in application turnaround, and enhances accuracy by up to 90%. The fully automated workflow minimizes manual intervention, allowing users to focus on decision-making through a streamlined graphical user interface (GUI), where they can effortlessly review and approve applications with greater confidence.- Used software
-
Docker Container-based application.
Backend: Python, FastAPI, pgSQL, Frontend: ReactJS, LLM stack: LangChain, meta-llama3.1 models.
- Difficulty of implementation
-
Complexity consists of:
- Variety of reinsurance application formats,
- Need for heavy prompt engineering when using small (5-10 billion parameters) LLMs.
- Project Description
-
The system provides end-to-end reinsurance automation. Reinsurance applications are collected through a special email box. Applications are processed automatically — the first scenario is to extract related values from the application (e.g. names, objects, premium, currencies) and to provide users with the key-value pairs in the GUI, which after user’s approval are imported to the decision-making system. While applications are being parsed, models also perform a few-shot analysis of underlying risks (e.g. how well the insured object is protected in case of fire).
The whole process of underwriting – application analysis, internal voting, correction of premium rates – can be performed in the system.
- Project geography
- Uzbekistan