MajjiKishore

I developed the Compliance AI Agent to tackle the complex and often manual process of product certification for international trade. My goal was to create a smart tool that could automatically check product documents against global regulations, making it easier for businesses to avoid costly rejections and delays.

How It Works: My Development Process

Building this project involved integrating a modern backend with a sophisticated AI pipeline. Here’s a breakdown of the key decisions and challenges I worked through.

Backend & Database

I chose FastAPI to build the core API. Its asynchronous capabilities were essential for creating a responsive user experience, especially for handling file uploads and processing in the background without making the user wait. For the database, I went with PostgreSQL because of its reliability. I also implemented Alembic to manage database schema changes, which allowed me to update the data structure smoothly as the project evolved.

The AI-Powered Pipeline

This was the most challenging and interesting part of the project. Getting from a raw user document to a final compliance score required several steps:

Document Processing: A major hurdle was accurately extracting text from various document formats like PDFs and scanned images. To solve this, I integrated Docling OCR into the pipeline.

Context-Aware Analysis: Simply extracting text wasn't enough. For the AI to understand the content, it needed context. I experimented with different methods and found that using Sentence Transformers to break the text into semantically meaningful chunks gave the best results. This step was crucial for improving the accuracy of the final analysis.

Core Logic: With the data cleaned and structured, I used Google's Gemini as the core reasoning engine. It analyzes the text chunks against regulatory rules, calculates a compliance score, and generates a clear, actionable report for the user.

Reference : https://github.com/majjikishore007/compliance_ai