DocMindAI

DocMind AI uses OCR and AI to automatically verify receipts, categorize expenses, and flag suspicious transactions, helping organizations prevent fraud and improve transparency.

Comment

Inspiration

Many NGOs, startups, and grant-funded organizations still manage expenses manually. This creates several problems:

Fake or duplicate receipts

Manual calculation errors

Delayed audits

Lack of financial transparency

Reduced trust from donors

For example:

An NGO receives ₹5,00,000 as a grant. If even 5% of expenses are fraudulent, that means:

Fraud amount = 5% of 5,00,000 = 0.05 × 5,00,000 = ₹25,000 loss

This inspired us to build DocMind AI, a system that automatically verifies expenses and detects suspicious transactions using AI.

What it does

DocMind AI is an AI-powered automated expense auditor.

It performs:

Receipt text extraction using OCR

Expense categorization using AI

Fraud detection

Anomaly detection

Explainable audit report generation

Example

Input receipt:

Vendor: ABC Travels

Amount: ₹18,500

Category: Travel

System checks:

Is the amount unusually high?

Is the receipt duplicated?

Is the vendor suspicious?

Output:

Status: Flagged

Reason: Amount exceeds average travel cost by 42%

How we built it

We designed DocMind AI as a multi-agent AI pipeline.

Each agent performs a specific task. Folder Structure: DocMindAI/ │ ├── backend/ │ ├── main.py │ ├── routes/ │ │ └── upload.py │ ├── services/ │ │ ├── ocr_service.py │ │ ├── categorization_agent.py │ │ ├── fraud_agent.py │ │ ├── anomaly_detector.py │ │ └── audit_agent.py │ ├── models/ │ │ └── isolation_forest.pkl │ └── database/ │ └── db_config.py │ ├── frontend/ │ ├── src/ │ │ ├── components/ │ │ │ ├── UploadForm.tsx │ │ │ ├── Dashboard.tsx │ │ │ └── ReportView.tsx │ │ ├── pages/ │ │ │ └── Home.tsx │ │ └── api/ │ │ └── api.ts │ └── package.json │ ├── data/ │ └── sample_receipts/ │ ├── requirements.txt └── README.md

Pipeline Steps

User uploads receipt

OCR extracts text

Categorization agent classifies expense

Fraud agent checks duplicates/inconsistencies

Anomaly detector finds abnormal patterns

Audit agent generates explanation

Flowchart of the System: User Uploads Receipt │ ▼ OCR Text Extraction │ ▼ Expense Categorization │ ▼ Fraud Detection Agent │ ▼ Anomaly Detection │ ▼ Audit Summary Agent │ ▼ Final Audit Report

Challenges we ran into

Extracting clean text from different PDF formats

Handling tables and numbers correctly

Ensuring accurate AI responses

Integrating AI processing with the backend

Optimizing performance for large documents

Extra Demonstrations (to impress the jury) Example 1: Invoice Calculation Product Price Quantity Pen ₹10 5 Notebook ₹50 3

Step-by-step calculation:

Pen total: 10 × 5 = ₹50

Notebook total: 50 × 3 = ₹150

Grand total: 50 + 150 = ₹200

Example 2: Average Expense Day Expense Day 1 ₹300 Day 2 ₹450 Day 3 ₹250

Total = 300 + 450 + 250 = ₹1000 Average = 1000 ÷ 3 = ₹333.33

Example 3: Percentage Calculation

If total budget = ₹5000 Used amount = ₹3500

Percentage used: (3500 ÷ 5000) × 100 = 70%

Remaining: 100 − 70 = 30%

Smart AI Response Example

User asks:

“How much budget is left?”

AI:

“30% of the budget remains, which equals ₹1500.”

Accomplishments that we're proud of

What we learned

Building DocMindAI gave us both technical and practical insights.

  1. Document Processing

We learned how different documents:

Contain structured data (tables, invoices)

Contain unstructured text (reports, articles)

Require different extraction techniques

We understood:

Text extraction from PDFs

Handling noisy or incomplete data

Converting raw text into structured information

  1. Natural Language Processing (NLP)

We learned how AI models:

Understand user questions

Extract relevant information

Generate summaries

Key concepts:

Tokenization

Context understanding

Prompt engineering

Information retrieval

  1. Machine Learning Concepts

We understood:

Classification models

Anomaly detection

Pattern recognition in data

Feature extraction from documents

  1. Backend–Frontend Integration

We learned how to:

Connect AI models with web interfaces

Handle file uploads

Process data in real time

Return results to the user

  1. Real-World Problem Solving

We realized:

Data is messy in real documents

AI must be accurate and reliable

Performance matters for large files

Simplicity improves user experience

What's next for DocMindAI

Short-Term Goals

Add image document support using OCR

Improve AI accuracy

Support more file types:

Word documents

Excel sheets

Add multilingual support

Mid-Term Goals

Smart financial analysis

Legal document summarization

Automatic report generation

Cloud storage integration

Long-Term Vision

DocMindAI will become:

A complete AI assistant that reads, understands, analyzes, and explains any document in seconds.

Future features:

Voice interaction with documents

Real-time collaboration

Enterprise-level document intelligence

Automated business insights

Anomaly Detection – Formulas and Example:

Anomaly detection identifies data points that differ significantly from normal patterns. Method 1: Z-Score Method Formula 𝑍=𝑋−𝜇/𝜎 ​ Where: X = data point μ = mean σ = standard deviation

If: ∣𝑍∣>3 → The data point is considered an anomaly.Example Sum

Dataset (expenses): 100, 120, 110, 115, 105, 500

Step 1: Calculate Mean μ=100+120+110+115+105+500/6 μ=1050/6=175

Step 2: Calculate Standard Deviation

σ=Square root(n∑(X−μ)^2) ​ (100−175)^2=5625 (120−175)^2=3025 (110−175)^2=4225 (115−175)^2=3600 (105−175)^2=4900 (500−175)^2=105625

Sum: 5625+3025+4225+3600+4900+105625=126000 𝜎=square root(126000)​ σ≈144.9

Step 3: Z-score for 500 Z=500−175/144.9 Z=144.9/325=2.24 If threshold = 2: → 500 is an anomaly Classification – Machine Learning Formula

Classification predicts a category or label.

Logistic Regression Formula Sigmoid Function P(y=1)=1/1+e^−z Example: Document Classification

Goal: Classify document as Normal (0) or Fraud (1)

Features: x1 = total amount

𝑥2 = number of transactions Given:

𝑤1=0.02 𝑤2=0.5 b=−4

Input:

𝑥1=2000 𝑥2=10 z=41 P=1/1+e^−41 P≈1 Result:

Since: P>0.5

Prediction: Fraud (Anomalous document)

Built With

Share this project:

Updates