Progress Reports

Task Breakdown

We used issue boards to keep track of  what to do , who's doing it and other things throughout the iteration of this project

Start to Demo 1:

  • Greg - Set up the frontend; Typescript react and create basic UI screens 
  • Naveen - Set up AWS account, ci/cd pipeline and work on PDF extraction pipeline
  • Jamey - Learn AWS; work on Log-in, Sign-up authentication end points
  • Nick - Set up AWS python CDK for future deployments; JWT introduction to Codebase

Demo 1 to Demo 2:

  • Greg - Created UI screens for open detail, submission intake, document viewer and other main tabs
  • Jamey - Worked on Email submission pipeline and assist with PDF extraction pipeline
  • Naveen- Worked on Document classification and assist with PDF extraction pipeline
  • Nick - Worked on Risk-assessment engine to be dynamic 

Demo 2 to Demo 3:

  • Greg - Finalized UI changes with client through constant feedback
  • Jamey - Finalized UI changes with risk-assessment; help with new integrations
  • Naveen - Continue to work towards pdf extraction accuracy and implement scoring engines
  • Nick - Integrate app scoring into risk-assessment engine

Post Demo 3:

  • Greg - Document Frontend
  • Jamey - Document Frontend/backend features worked on
  • Naveen - Attempt OCR approaches; document
  • Nick - document risk-assessment
  • As a group: hand off to client

Stage 1 (Demo 2) Report

The team has completed core features such as:

  • user authentication, 
  • session management, 
  • frontend-backend integration 
  • functional PDF upload and processing pipeline.

The system can already take uploaded PDFs and generate structured JSON data using AI-powered extraction tools. The platform is deployed on AWS and an automated CI/CD pipelines are in place to support faster development and testing. Current work focuses on improving performance, expanding frontend pages and features, and building additional data processing capabilities.

Stage 2 (Demo 2) Report 

The team added features such as a:

  • document viewer
  • user settings management 
  • detailed data pages
  • RAG-powered chatbot for asking questions about extracted insurance data.

On the backend, we improved PDF extraction accuracy, built a risk assessment engine, added email-based submission workflows, and created document classification systems. The platform also introduced observability tools for tracing extraction workflows and developed a custom stateless, graph-based system for managing AI outputs dynamically and efficiently. Current work focuses on connecting the risk engine to the main application, improving pdf extraction accuracy, and testing with real-world data.

Final Report: 

ug_sb_1_ai-insurance-underwriter_perpendo-FINAL-REPORT.pdf

Meeting Minutes

Our meetings consisted of client meetings and TA meetings. We met with our client twice a week to discuss any updates and bi-weekly with our ta to share our progress so far in the application. 

  • Meet with our client twice a week (Tuesday 5:30 PM & Friday 11:00 AM/11:30 AM)
  • Meet with TA bi-weekly (Friday 11:00 AM)

Below is a rough sketch of what was developed and discussed through the weeks

Weeks 1–3: Planning & Core Architecture

Alot of the meetings during this time period was just getting started and deciding basic things like the tech stack, environment and etc. Nick proposed a AWS set up that follows a serverless type approach. We ended up picking for the frontend: Typescript react, backend: python, and cloud: AWS. Our client liked the set up.

  • Designed frontend and backend architecture
  • Set up AWS infrastructure using Lambda, S3, and DynamoDB
  • Created authentication flow with login, signup, and JWT sessions
  • Built initial React + TypeScript frontend structure

Weeks 4–7: File Upload & Processing Pipeline

Alot of features needed to get added in order to keep up with pace of the project. We agreed to add in the smaller features of the application quickly (pdf/xlsx uploads), and basic UI screens. Once those were done the focus was shifted to the pdf extraction pipeline and the document classification needed for it.

  • Implemented PDF/XLSX upload workflows
  • Connected uploads to S3 storage and backend processing
  • Built document classification system for underwriting files
  • Developed extraction pipeline using AI agents 
  • Stored extracted structured data in DynamoDB
  • Built dashboard and submission intake pages
  • Developed Open Detail page with organized underwriting tabs
  • Added document viewer with secure file rendering
  • Improved UI navigation and reusable shared components

Weeks 8–10: UI Importance and Risk Introduction  

After the heavy backend weeks, frontend needed an expansion. Greg was our main frontend guy, but Jamey and Naveen provided some assistance with the frontend UI leading up to demo 2. The goal was to make it clean according to the client's UI templates. Nick worked heavily on the risk-assessment engine in the backend. It was discussed several times over call about needing to be deterministic to the rules.

  • Added RAG-powered chatbot for underwriting Q&A
  • Developed scoring engines for risk analysis
  • Built intelligent document-specific extraction agents
  • Introduced semantic-cast graph system for predictable AI workflows
  • Added observability dashboards using Open-Telemetry

Weeks 11–13: Observability & Admin Tools

Main things talked about here was PDF extraction accuracy. While it got some things right, it got things wrong as well. Talks about how to improve it where discussed. Role-based access was also another conversation that happened. The client also asked to implement a separate scoring engine for MVR to compare with the pdf extraction accuracy.

  • Created admin screens
  • Continued approaches to improve PDF extraction accuracy
  • Appetite scoring integration to risk assessment
  • MVR scoring engine & Loss history scoring engine introduced
  • Finalized a-lot of features

Weeks 14–15: Last Touches and Handoff

The last thing our client wanted was to change the UI to his new UI templates which greg and Jamey took the task of doing. He wanted to try and OCR approach to see how it would work, so Naveen took that task and all of us started to discuss documentation and the handoff. We meet in-person to do the hand off so our client would not be confused.

  • UI changes finalized
  • MVR scoring engine & loss history engines added
  • Hand Off to client (copied repo)
  • Helped Client set up pipelines
  • Documentation