Case Study MIA
Building an AI mortgage assistant that extracts structured financial data and answers complex guideline questions.
MIA (Mortgage Intelligence Assistant) is an AI-powered mortgage operations platform built for Greenbox Loans. It enables underwriters, loan officers, and processors to query mortgage guidelines, process statements, and extract transaction data instantly.
Overview
Supercharging mortgage operations with LLMs and structured parsing engines.
MIA combines a smart conversational AI chat assistant with internal administrative tools, document indexing, role-based access control, and automated PDF statement extraction workflows.
I contributed to building and scaling MIA's core components: multiple chat modes, PDF document knowledge systems, and backend extraction logic. To handle heavy computation without compromising UI responsiveness, we offloaded tasks like PDF extraction to background workers with real-time status monitoring and notifications.
Collaboration & Engineering Practices
We worked in an Agile setup, collaborating closely with cross-functional product and data engineering teams to ship AI features in short, iterative development cycles.
Due to the complexity of parsing financial and loan documents, focus was kept on high reliability, structured schema validation, and clean handoffs between the frontend chat interfaces, RAG data ingestion pipelines, and background processing systems. I also implemented detailed environment configuration checks to guarantee smooth deployments.
Core Responsibilities
- Built a robust document knowledge base where admins can manage competitor program details, provider indexes, uploaded PDFs, and searchable mortgage files.
- Implemented persistent chat history, conversation URLs, and role-based, permission-aware chat access.
- Built a smart bank statement extractor that uploads PDF statements, processes them with AI, extracts credits/debits, validates transaction totals, and saves results.
- Offloaded heavy PDF processing into resilient background jobs complete with retry logic, concurrency limits, artifact storage, and email notifications.
- Designed and integrated role-based access controls for Admin, Operations, and Sales roles.
- Created administrative tools for user permissions management, Excel template matching, saved extractions history, and secure CSV/Excel data export.
- Implemented structured environment validation, Supabase session handling, storage integrations, and optimized database models.
Key Accomplishment
By moving heavy PDF statement analysis into queue-based background workers with automatic retries and concurrent task execution, we successfully reduced average UI request timeouts to 0% and ensured users always receive email notifications and persistent database updates when extraction runs finish.