AuditIQ — AI-Powered Invoice Audit System

Built for the Mosaic Wellness Fellowship · Finance Operations Track

AuditIQ is an AI-powered invoice verification system that automatically detects overcharges, GST mismatches, duplicate invoices, and calculation errors across any invoice format — saving finance teams hours of manual work every billing cycle.

🔗 Live Demo → auditiq-zmffcnkmvp5g2ybxhr6ccp.streamlit.app

---

The Problem

5–10% of vendor invoices contain overcharges. For a D2C brand spending ₹50L/month on logistics and raw materials, that's ₹2.5–5L lost every single month — silently, because manual invoice checking is slow, error-prone, and nobody has time to do it properly.

AuditIQ solves this by automating the entire audit pipeline in minutes.

---

What It Does

Upload any invoice. AuditIQ extracts every field using AI, cross-checks it against your contract rates, and flags every discrepancy with a clear explanation and exact overcharge amount.

Check	What It Catches
Rate Overcharge	Line item rates higher than signed contract
GST Verification	GST % doesn't match contracted rate
HSN Code Validation	GST rate contradicts GST Council database (80+ HSN codes)
Mystery Surcharges	Line items not present in the contract at all
Calculation Errors	Arithmetic mistakes — qty × rate and subtotal + GST
Duplicate Detection	Same invoice resubmitted in a batch or across history
Historical Deviation	Prices that have crept up >10% from vendor's historical average

---

Supported Formats

Works on any invoice format without pre-configuration:

PDF Scanned PDF PNG JPG WEBP TIFF BMP Excel (.xlsx/.xls) Word (.docx)

---

Key Features

🔍 AI Extraction

• Extracts vendor name, invoice number, date, line items, quantities, rates, GST, HSN codes
• Uses vision model for scanned/image invoices, text model for digital PDFs
• 3-library fallback for scanned PDFs (PyMuPDF → pdf2image → pypdfium2)
• 4-stage JSON recovery — never crashes on malformed AI output

📊 Accuracy Engine

• Weighted confidence scoring — arithmetic checks (99%) carry more weight than heuristic checks (88%)
• Grades every audit A+ / A / B / C / D
• Flags weak extraction fields before you rely on results
• GST double-flag deduplication — same root cause counted once

💼 Finance Dashboard

• Total billed vs correct amount
• Payment hold alerts for HIGH severity issues
• Vendor risk scoring across audit history
• Downloadable Word report with findings and recommended actions

🗂️ Rate Card Management

• Upload rate cards in any format — PDF, Excel, image, CSV
• AI extracts vendor name, item rates, and GST automatically
• Vendor database built entirely from your own contracts

---

Tech Stack

Component	Technology
UI	Streamlit
LLM Inference	Groq (llama-3.3-70b + llama-4-scout-17b)
PDF Extraction	pdfplumber + PyMuPDF
Image Processing	Pillow
Report Generation	python-docx
Deployment	Streamlit Cloud

---

Project Structure

AuditIQ/
├── app.py                  # Streamlit UI — 6 tabs
├── extractor.py            # AI invoice extraction
├── auditor.py              # All 7 audit checks + HSN database
├── accuracy_engine.py      # Weighted confidence scoring
├── history_manager.py      # Audit history + pattern detection
├── rate_card_manager.py    # AI rate card extraction
├── report_generator.py     # Word .docx report (pure Python)
├── requirements.txt
├── .streamlit/
│   └── config.toml
└── data/
    ├── audit_history.json
    └── custom_rate_cards.json

---

Running Locally

# 1. Clone the repo
git clone https://github.com/harman123-cloud/AuditIQ.git
cd AuditIQ

# 2. Install dependencies
pip install -r requirements.txt

# 3. Add your Groq API key
echo "GROQ_API_KEY=your_key_here" > .env

# 4. Run
python -m streamlit run app.py

Get a free Groq API key at console.groq.com

---

Test Invoices

Six test invoices are included to validate every detection pattern:

File	Pattern	Expected Flag
01_rawmat_rate_overcharge.pdf	Steel ₹4,500→₹5,200, Aluminum ₹3,200→₹3,900	Rate Overcharge × 2
02_fastship_gst_wrong.pdf	GST 18%→28% + unlisted surcharge	Incorrect GST Rate + Mystery Surcharge
03_brandboost_calc_error.pdf	3×₹15,000 billed as ₹51,000	Line Item Calc Error + Calc Error
04_rawmat_DUPLICATE.pdf	Same invoice number as #01	Duplicate Invoice
05_rawmat_hsn_gst_mismatch.pdf	HSN 4819 (12%) billed at 18%	HSN Code GST Mismatch
06_brandboost_CLEAN.pdf	Everything correct	No flags — Grade A+

---

Built With

• Streamlit
• Groq
• pdfplumber
• PyMuPDF
• python-docx
• Pillow

---

Built using Claude Code as part of the Mosaic Wellness Fellowship application.