DocMindAI

• 

  Topic of our Project

• 

  Problem description

• 

  Solution

• 

  Tech Stack Used

• 

  Workflow of our project

• 

  Login based Users

• 

  Admin Overview

• 

  Anomaly Detection

• 

  Benefits and Future Enhancements

Inspiration

Many NGOs and small organizations still rely on manual expense tracking using paper receipts and spreadsheets

Manual processes lead to human errors, duplicate claims, delayed audits, and lack of transparency

Donors and stakeholders expect accountability, but affordable audit tools are often unavailable

We wanted to use AI to reduce manual effort and support transparent, trustworthy financial management

DocMindAI was inspired by the need for an explainable AI system that assists audits rather than replacing human judgment

What it does

Automatically extracts information from receipts and invoices

Understands expense context such as vendor, amount, date, and category

Verifies expenses by checking for duplicates, inconsistencies, and abnormal spending

Detects suspicious transactions using anomaly detection

Provides clear, human-readable explanations for every audit decision

Displays results in an easy-to-use dashboard for quick review

How we built it

Used PaddleOCR-VL to extract text and layout information from receipt images

Built a multi-agent system powered by ERNIE 4.5 for semantic reasoning

Assigned each agent a clear role:

Document understanding

Expense categorization

Fraud and consistency checking

Audit explanation generation

Integrated Isolation Forest to statistically detect anomalous spending patterns

Implemented the backend in Python and connected it to a visualization dashboard

Challenges we ran into

Handling diverse receipt formats with varying quality and layouts

Extracting accurate totals from noisy or poorly scanned documents

Designing agent responsibilities without overlap

Balancing accuracy, explainability, and performance within a hackathon timeline

Ensuring outputs were understandable for non-technical users

Accomplishments that we're proud of

Built a complete end-to-end AI expense auditing system

Successfully integrated PaddleOCR-VL with ERNIE in a multi-agent architecture

Created an explainable system that shows why an expense is flagged

Demonstrated real-world applicability for NGOs and small organizations

Aligned the project with transparency, accountability, and social impact

What we learned

Real-world financial documents are messy and inconsistent

Deep-learning OCR performs much better than traditional OCR methods

Multi-agent systems improve modularity and explainability

Combining statistical models with LLM reasoning leads to more reliable decisions

Explainability is essential for trust in financial AI systems

What's next for DocMindAI

Fine-tune PaddleOCR-VL for handwritten and multilingual receipts

Further fine-tune ERNIE for financial auditing and compliance reasoning

Add real-time alerts for suspicious transactions

Generate automated audit and donor-ready reports

Enable offline and edge deployment for low-connectivity environments

Built With

• camel-ai
• ernie
• paddleocr-vl
• pandas
• python
• scikit-learn