⚠️ Note: This repository contains the original Forage job simulation (Tasks 1–4). It served as the foundation for a much more advanced system. For the full Agentic AI build — with reasoning, memory, ethical guardrails, and Streamlit deployment — see 👉 CreditRiskAI-Agentic-Collections
Role: AI Transformation Consultant
Client: Geldium Finance
Program: GenAI Powered Data Analytics Job Simulation by Tata iQ
This project simulates my work as an AI Transformation Consultant at Tata iQ, helping Geldium Finance — a digital lending company — reduce its high credit card delinquency rate using GenAI and advanced analytics.
The goal was to move beyond traditional manual collections by building an intelligent, ethical, and explainable AI solution that identifies at-risk customers early and recommends the most effective intervention strategies.
- Perform Exploratory Data Analysis (EDA) and improve data quality
- Build a predictive model to forecast customer delinquency
- Translate model insights into actionable business recommendations
- Design a high-level AI-powered autonomous collections system with strong ethical guardrails
Tata-iQ-GenAI-Data-Analytics-Forage/ ├── README.md ← Main project documentation ├── LICENSE ← Optional (MIT or Apache) ├── requirements.txt ← List of Python packages ├── .gitignore ← Ignore unnecessary files │ ├── data/ │ ├── raw/ │ │ └── Delinquency_prediction_dataset.xlsx ← Original file │ └── processed/ │ ├── Cleaned_Delinquency_Dataset.xlsx │ └── Final_Dataset_with_New_Target.xlsx │ ├── notebooks/ │ ├── 01_EDA_Task1.ipynb │ ├── 02_Predictive_Modeling_Task2.ipynb │ ├── 03_Business_Recommendations_Task3.ipynb │ └── 04_Agent_Prototype.ipynb ← (I will add this later) │ ├── reports/ │ ├── Task1_EDA_Report.pdf │ ├── Task2_Model_Report.pdf │ └── Task3_Business_Recommendations.pdf │ ├── visuals/ │ ├── delinquency_boxplots_improved.png │ ├── feature_importance.png │ └── confusion_matrix.png │ └── presentation/ └── Task4_Final_Presentation.pptx
- Programming Language: Python 3
- Data Analysis: Pandas, NumPy
- Visualization: Matplotlib, Seaborn
- Machine Learning: Scikit-learn (Decision Tree Classifier)
- Environment: Jupyter Notebook
- Other Tools: GenAI (Grok, ChatGBT) for code generation, insights, and report structuring
- Cleaned inconsistent data and handled missing values using group median imputation
- Identified
Missed_PaymentsandCredit_Utilizationas the strongest risk indicators - Created an improved target variable (
Delinquent_Account_New) for better model learning
- Built a Decision Tree model achieving:
- Accuracy: 83%
- AUC-ROC: 0.868
- F1-Score: ~0.83 (balanced)
Missed_Paymentscontributed 62.17% to model decisions
- Developed a SMART intervention strategy: Targeted SMS + flexible payment plan pilot for high-risk customers
- Goal: Reduce 30+ day delinquency by 15% in the pilot group
- Designed a conceptual Agentic AI Collections System with:
- Real-time risk scoring
- Autonomous decision-making with human oversight
- Strong ethical guardrails (fairness, explainability, compliance)
- The importance of data quality and thoughtful feature engineering in financial modeling
- How to improve weak target variables using domain logic
- The critical balance between model performance and explainability in regulated industries
- How to translate technical insights into clear business recommendations
- The value of Agentic AI and responsible AI design (guardrails, bias detection, human-in-the-loop)
This simulation gave me hands-on experience in the full lifecycle of an AI project in financial services — from raw data to strategic automation.
I'm actively seeking opportunities in Dtata Science, Data Analytics, AI/ML, and Financial Services.
Feel free to connect if you're working on similar projects or have any feedback!