Attention seeker

Project: LoanIQ

Team Name: Attention Seekers

Inspiration

We were frustrated by the archaic manual processes in the multi-trillion dollar loan market, where bankers still spend hours manually "spreading" data from static PDFs. We wanted to build an "always-on" junior analyst that could transform these dead documents into dynamic, queryable insights, merging financial rigor with modern, accessible design.

What it does

LoanIQ allows users to upload PDF loan agreements and instantly visualizes critical risks by extracting metadata into a structured, glassmorphic dashboard. It acts as an intelligent assistant, offering audio briefings of legal terms and automatically generating PowerPoint presentations for executive review. The system bridges the gap between unstructured legalese and structured financial data using a hybrid AI.

How we built it

We built a responsive UI using Gradio and Python, integrated with OpenAI's GPT-4 for semantic analysis and PyMuPDF for text extraction. The backend orchestrates a hybrid pipeline where regex handles precise values while the LLM captures nuanced legal clauses. We added multimodal capabilities using OpenAI's TTS for audio summaries and python-pptx for automated slide generation.

Challenges we ran into

Getting the LLM to consistently output valid JSON from messy legal documents was a major hurdle requiring extensive prompt engineering and regex fallbacks. Managing the application state across multiple tabs in Gradio while handling real-time processing proved complex. We also had to balance token limits with the massive size of typical 200+ page loan agreements.

Accomplishments that we're proud of

We successfully created a "hybrid extraction engine" that marries the flexibility of Generative AI with the precision of traditional regex. We delivered a polished, "orange-themed" UI that feels premium and modern, breaking the mold of boring financial software. The "one-click PPT" feature was a standout win, instantly converting raw data into a client-ready presentation.

What we learned

We learned that while LLMs are powerful, strict schema enforcement is essential for reliable enterprise tools. We discovered that providing multimodal outputs (like audio and slides) drastically improves the usability of legal tech. We also found that Gradio is capable of powering complex, stateful applications when architected correctly.

What's next for Attention seeker

We plan to implement a RAG (Retrieval-Augmented Generation) system to allow users to "chat" directly with their loan portfolio. We aim to scale the backend to support simultaneous cross-document comparison for identifying market trends. Finally, we will deploy the containerized application to a serverless cloud environment.