A two-sided fintech intelligence platform — a behavioral nudge engine for users, and a monetization analytics layer for banks.
SmartTax models how financial institutions can extract actionable signals from customer savings behavior: Prophet forecasts category-level spend, KMeans segments risk appetite, and Power BI surfaces partnership triggers for travel/telecom tie-ups and pre-approved loan targeting.
 Spending Analytics |
 Customer Segmentation |
 Goal Analytics |
- What SmartTax Actually Is
- The Problem Worth Solving
- Platform Architecture
- Technology Stack
- Data Engineering Pipeline
- Machine Learning Pipeline
- API Reference
- Business Intelligence Layer
- Strategic Value for Banks
- Project Structure
- Getting Started
- Future Enhancements
- Contributors
SmartTax is not a savings app. It is a dual-stakeholder fintech platform designed around one core insight:
Every UPI transaction is a behavioral signal. Most banks ignore it. SmartTax doesn't.
| Stakeholder | Engine | What It Does |
|---|---|---|
| User | Behavioral Nudge Engine | Goal creation triggers ML → Prophet forecasts spend → micro-tax % is calculated per category → user is nudged to save with every transaction |
| Bank | BI Intelligence Layer | KMeans segments customers by risk appetite → Power BI surfaces partnership triggers, churn signals, and pre-approved loan candidates |
The user-facing side solves financial discipline. The bank-facing side solves customer monetization. Both are powered by the same underlying data pipeline — which is the architectural bet that makes this interesting.
For users:
Most people don't fail to save because they lack intention. They fail because the gap between deciding to save and actually saving is never bridged. There's no mechanism that converts a ₹500 Zomato order into ₹30 toward a laptop fund — automatically, contextually, without friction.
For banks:
A customer completes 200 UPI transactions. The bank records them. That's it. No behavioral segmentation. No spend forecasting. No signal that this customer is 3 months away from completing a ₹50,000 savings goal — and therefore primed for a fixed deposit upsell. The data exists. The intelligence doesn't.
SmartTax bridges both gaps.
┌──────────────────────────────────────────────────────────┐
│ DATA LAYER │
│ Kaggle Banking CSV → ETL (Python/pandas) │
│ → MongoDB Atlas │
│ Collections: users, transactions, categories, │
│ goals, ml_percentages │
└──────────────────────────────────────────────────────────┘
↓
┌──────────────────────────────────────────────────────────┐
│ ML LAYER │
│ Prophet → time-series spend forecasting per category │
│ KMeans → risk-appetite segmentation (BI only) │
│ Tax Formula → category-wise % required to hit goal │
│ Output → ml_percentages collection │
└──────────────────────────────────────────────────────────┘
↓
┌──────────────────────────────────────────────────────────┐
│ BACKEND LAYER │
│ Flask (Python) │
│ → Auth, Goals, Transactions │
│ → Triggers ML pipeline on goal create/update │
│ → Serves /api/insights endpoints for Power BI │
└──────────────────────────────────────────────────────────┘
↓ ↓
┌────────────────────┐ ┌────────────────────────────┐
│ USER SYSTEM │ │ BANK SYSTEM │
│ React + Vite │ │ Power BI Desktop │
│ ───────────── │ │ ──────────────── │
│ Login / Register │ │ Pulls from /api/insights │
│ Goal Dashboard │ │ Segmentation Reports │
│ Transaction Feed │ │ Goal Completion Analytics │
│ Tax Recommendations│ │ Partnership Trigger Signals│
└────────────────────┘ └────────────────────────────┘
| Layer | Technology | Why This Choice |
|---|---|---|
| Dataset | Kaggle Banking Dataset | Real transaction distribution patterns |
| ETL | Python, Pandas | Flexible transformation with audit trail |
| Database | MongoDB Atlas | Schema-flexible for evolving goal/category structures |
| ML Forecasting | Facebook Prophet | Handles seasonality and missing data without fine-tuning |
| User Segmentation | KMeans Clustering | Interpretable clusters map cleanly to BFSI segments |
| Backend | Flask | Lightweight REST layer; triggers ML on goal events |
| Frontend | React + Vite | Fast iteration, component-driven dashboard |
| BI | Power BI | Enterprise-grade reporting; REST connector to Flask |
Raw Kaggle banking transactions are cleaned, normalized, and enriched through:
- Data cleaning and null handling
- Merchant category standardization
- User-level aggregation
- Synthetic user generation to simulate a realistic customer base
Pipeline output:
50 synthetic users
7,601 transactions
Loaded into MongoDB Atlas
Customers are clustered into three behavioral segments using unsupervised learning across five behavioral features:
| Feature | Signal |
|---|---|
| Total spend | Absolute consumption level |
| Average transaction size | Ticket size preference |
| Transaction count | Engagement frequency |
| Spending diversity | Category spread |
| Spending variability | Behavioral consistency |
| Segment | Characteristics | Bank Opportunity |
|---|---|---|
| Conservative | Low spend, stable patterns | Savings products, FDs |
| Moderate | Balanced behavior | Cross-sell insurance, SIPs |
| Aggressive | High spend, high variability | Premium cards, pre-approved loans |
When a user sets a savings goal, the ML pipeline fires automatically:
Input: Goal → "Laptop Fund | ₹50,000 | 12 months"
Output: Category-wise tax percentages stored in ml_percentages
Prophet generates next-month spend predictions per category using the user's historical transaction data:
Food → ₹4,500 predicted
Shopping → ₹2,000 predicted
Travel → ₹1,500 predicted
Healthcare → ₹1,000 predicted
Prophet is particularly well-suited here: it handles irregular UPI spending patterns, accommodates weekday/weekend seasonality, and doesn't require clean, gapless time series — exactly the conditions real banking data presents.
SmartTax computes the category-wise contribution rate needed to hit the goal within the target duration:
| Category | Predicted Spend | Recommended Tax | Monthly Contribution |
|---|---|---|---|
| Food | ₹4,500 | 6% | ₹270 |
| Shopping | ₹2,000 | 12% | ₹240 |
| Travel | ₹1,500 | 10% | ₹150 |
The percentages are inversely weighted — categories with elastic demand (Shopping, Travel) absorb higher tax rates than necessities (Food, Healthcare).
Transaction: Food ₹500
Applied Tax: 6%
Diverted: ₹30 → Laptop Fund
The user can accept, modify, or override any recommendation. Accepted amounts are directly credited to goal progress.
| Attribute | Detail |
|---|---|
| Purpose | Category-level monthly spend prediction |
| Input | Historical transactions, goal target, duration |
| Output | Predicted spend per category, tax % per category, goal feasibility score |
| Why Prophet | Designed for business time series with trend shifts and seasonality; performs well on sparse data without hyperparameter tuning |
| Attribute | Detail |
|---|---|
| Purpose | Customer behavioral profiling for bank BI |
| Input | Aggregated user transaction features |
| Output | Conservative / Moderate / Aggressive segments |
| Why KMeans | Produces interpretable, discrete segments that map directly to BFSI product targeting logic |
POST /api/auth/register
POST /api/auth/loginGET /api/goals/<customer_id>
POST /api/goals # Triggers ML pipeline
PUT /api/goals/<goal_id> # Triggers ML recalculationPOST /api/transactions
GET /api/transactions/<customer_id>GET /api/percentages/<customer_id>/<goal_id>
PUT /api/percentages/<customer_id>/<goal_id>GET /api/insights/spending # Category-level spend aggregates
GET /api/insights/goals # Goal completion rates, on-track vs off-track
GET /api/insights/users # Segment distribution, behavioral profiles
GET /api/insights/overrides # ML recommendation override patternsSmartTax's bank-facing layer is not a dashboard bolted on at the end. It is a first-class product surface — a REST-fed Power BI reporting layer that converts behavioral transaction data into actionable bank intelligence.
Intelligence surfaced:
- Highest-volume and highest-value transaction categories
- Category-level seasonality and trend signals
- Per-customer spending velocity
Partnership triggers this unlocks: Travel spike detected → co-branded travel card offer Electronics category dominant → EMI financing campaign Food delivery dominant → cashback partnership with aggregators
Intelligence surfaced:
- Segment distribution across the customer base
- Spending intensity by cluster
- Behavioral drift (customers moving between segments)
Targeting logic this unlocks: Aggressive segment → pre-approved personal loan or credit limit increase Conservative segment → FD/RD nudge campaigns Moderate segment → SIP or insurance cross-sell
Intelligence surfaced:
- Goal completion rates by category (Travel, Education, Emergency, etc.)
- On-track vs off-track customer distribution
- Override patterns (where users reject ML recommendations)
Retention and product logic this unlocks: Off-track users → savings nudge or automated RD offer Goal completion approaching → upsell to next savings product High override rate → signals distrust; candidate for financial advisory outreach
SmartTax answers a question most banks don't know how to ask:
What does a customer's UPI history tell me about what to sell them next — and when?
| Use Case | Signal | Action |
|---|---|---|
| Customer Retention | Goal progress stalling | Proactive savings nudge or RD offer |
| Loan Targeting | Off-track goals + high spend variability | Pre-approved personal/education loan |
| Product Cross-sell | Goal completion near | FD, next-tier savings account |
| Partnership Triggers | Category spend spikes | Co-branded card or cashback tie-up |
| Churn Prevention | Declining transaction frequency | Re-engagement campaign |
| Credit Risk Scoring | Segment + savings consistency | Input signal for credit decisioning |
SmartTax/
│
├── etl/
│ └── etl.py # Kaggle → pandas → MongoDB pipeline
│
├── ml/
│ ├── ml.py # Prophet forecasting + tax % calculation
│ └── cluster.py # KMeans segmentation
│
├── backend/
│ ├── routes/ # Auth, Goals, Transactions, Insights
│ ├── models/ # MongoDB schemas
│ └── app.py # Flask application entry point
│
├── frontend/ # React + Vite user dashboard
│
├── powerbi/
│ └── SmartTax.pbix # Power BI report (REST-connected to Flask)
│
└── README.md
# Clone the repository
git clone https://github.com/Dev-06-06/SmartTax.git
cd SmartTax
# Backend setup
cd backend
pip install -r requirements.txt
python app.py
# Frontend setup
cd ../frontend
npm install
npm run dev
# Run ETL pipeline
cd ../etl
python etl.py
# Run ML pipeline
cd ../ml
python ml.py # Prophet forecasting
python cluster.py # KMeans segmentationBefore running: configure your MongoDB Atlas connection string and required environment variables in the backend
.envfile.
| Enhancement | Strategic Value |
|---|---|
| Real UPI Integration | Live behavioral data replaces synthetic dataset |
| Open Banking APIs | Multi-bank transaction aggregation |
| AI Financial Assistant | Conversational goal planning layer |
| Fraud Detection | Anomaly flagging on transaction patterns |
| Credit Risk Scoring | Savings behavior as a creditworthiness signal |
| Real-Time Streaming | Kafka-based pipeline for live BI updates |
| Recommendation Engine | Investment product matching by segment + goal |
Dev B.Tech Computer Science & Engineering FinTech · Data Engineering · Machine Learning · Business Intelligence