• 

  Perforated AI at work

Bank Lead Scoring with Dendritic Optimization

Intro

Description:

This project demonstrates the application of Dendritic Optimization to a real-world banking lead scoring problem. Global banks process millions of lead calls daily, and inefficient targeting wastes agent time and expensive server costs. We built a Lead Scoring Engine to predict which customers are most likely to accept a term deposit offer, optimized for deployment on edge devices like bank agent tablets.

Project Impact

Standard deep learning models for tabular data in banking are often over-parameterized (700,000+ parameters), making them too slow for low-power edge devices and too expensive to run on cloud infrastructure for millions of transactions. An optimized lead scoring model matters because it enables real-time AI inference on agent tablets, reducing operational costs by an estimated 40% through better lead prioritization. This allows banks to focus resources on high-value prospects while maintaining prediction quality, eliminating the need for expensive cloud-based inference and enabling zero-latency decisions at the point of customer contact.

Usage Instructions

Installation:

pip install pandas torch scikit-learn wandb perforated-ai

Run:

1. Generate the dataset:

   python setup_data.py
   

2. Run the dendritic optimization training:

   python train.py --use_dendritic 1
   

3. Build deployable models:

   python build_demo.py
   

4. Run the demo application:

   python run_demo.py
   

Results

This bank lead scoring project demonstrates that Dendritic Optimization can achieve massive parameter reduction while retaining business-critical accuracy. Comparing the standard baseline model to the dendritic optimized model:

Compression Results:

• Percent Parameter Reduction: 81%
• Performance Retention: 98% (only 1.5 percentage point accuracy drop)
• Deployment Impact: Model size reduction enables edge deployment on tablets/ATMs with zero-lag inference

The optimizer identified that 81% of the baseline model's capacity was redundant, as shown in PAI/PAI_beforeSwitch_128best_test_scores.csv. We achieved comparable business value with 1/5th the original size, making the model deployable on low-power edge devices.

Weights and Biases Sweep Report

View Full Interactive W&B Report

The W&B report demonstrates comprehensive experimentation with the dendritic optimization process, showing how the architecture search dynamically discovered the optimal 135k-parameter model structure.

Additional Files

Code Structure:

• setup_data.py - Generates the banking dataset
• train.py - Main training script with dendritic optimization
• build_demo.py - Reconstructs the optimized architecture in pure PyTorch
• run_demo.py - Demo application for lead prioritization

Zero-Dependency Deployment:

This project implements a "Factory Pattern" for production deployment:

1. Search Phase: Used train.py with Perforated AI to discover the optimal 135k-parameter architecture
2. Build Phase: Reconstructed this specific shape in pure PyTorch (build_demo.py)
3. Deploy Phase: The resulting optimized_model.pth runs on any standard device without requiring the Perforated AI library installed

This optimized model is being integrated into a proprietary Marketing Intelligence Tool to automate lead prioritization for field agents, with a full presentation planned for the Hack2Skill Buildathon on January 25, 2026.

Built With

• perforatedai
• python
• pytorch
• scikit-learn
• wandb