AI-Driven Financial Insight Pipeline

Project Overview

I developed a comprehensive AI-powered pipeline to streamline financial data integration and enhance revenue forecasting for a business that manufactures and sells number plates across India. The primary challenge was consolidating internal sales data with external transaction records from a payment gateway (PhonePe), which initially only provided PDF exports.

Key Components and Workflow

1. Data Ingestion and Processing

• Internal data sources: invoices, transactions, and order details stored in SQL databases.

• External data: PhonePe payment gateway, which only provided transaction records as PDF exports.

• To process these PDFs, I leveraged NotebookLM, which extracted and structured the raw transaction data into clean, usable tables. This bridged the gap between unstructured files and structured financial datasets.

  
  

2. Pipeline Orchestration with Airflow

• Implemented Apache Airflow to automate the end-to-end workflow:

  
  

  1. Retrieve raw files (PDFs/CSVs).

  2. Run data extraction via NotebookLM.

  3. Load structured data into SQL for storage.

  4. Apply predictive models for forecasting.

  
  

• Airflow ensured the pipeline ran reliably on a schedule with logging and monitoring.

  
  

3. Predictive Modeling with AI

• Leveraged Gemini and GPT-based models to forecast revenue using both historical sales and payment gateway data.

• Incorporated external macroeconomic signals (e.g., seasonal demand, regional sales variation) suggested by AI to enrich model accuracy.

• AI also auto-documented assumptions in natural language, enabling business users to understand forecasts without needing technical details.

  
  

4. Automated Variance Analysis and Insights

• Designed the pipeline to run variance analysis between forecasts and actuals automatically.

• Used AI to generate narrative insights explaining variances

• Delivered results through interactive, queryable reports so executives could ask “what-if” questions.

  
  

Outcomes and Impact

• +20% improvement in revenue forecast accuracy.

• Identified three top-performing product lines, enabling targeted marketing and boosting sales revenue by 15% in six months.

• Reduced manual reporting effort by automating data extraction and variance analysis, saving the finance team 10+ hours per week.

• Delivered board-ready executive summaries and dashboards that updated automatically as new data arrived.

Technologies Used

• Python, SQL, Apache Airflow

• Gemini, GPT, NotebookLM

• PDF data extraction, predictive modeling, automated reporting