Schneider Electric - Quick2

https://datathon.quick2query.com/

Inspiration

Sales teams at Schneider Electric face a critical challenge: understanding why a machine learning model predicts whether an opportunity will be won or lost. While predictive models can forecast outcomes, the lack of transparency makes it difficult for sales professionals to trust and act on these predictions. We were inspired to bridge this gap between AI predictions and human decision-making by creating an explainability platform that transforms complex SHAP values into clear, actionable insights.

What it does

Quick2 is an AI-powered explainability dashboard that provides transparent, human-readable explanations for sales opportunity predictions. The platform:

• Predicts sales opportunity outcomes (WON/LOST) using a CatBoost classifier with 70%+ F1 score
• Explains each prediction using SHAP (SHapley Additive exPlanations) values to show exactly which features influenced the model's decision
• Visualizes feature importance through interactive charts, confusion matrices, and performance metrics
• Generates natural language explanations via LLM integration, translating technical SHAP values into business-friendly narratives
• Analyzes model performance across different segments (competitor presence, customer history, product mix)
• Filters predictions by confidence level, accuracy, and outcome to help sales teams prioritize their efforts

How we built it

We built Quick2 using a modern ML stack optimized for explainability:

• Model: CatBoost classifier trained on historical sales data with features including customer interaction history, product mix, competitor presence, and regional factors
• Explainability: SHAP library for calculating feature contributions to individual predictions and global feature importance
• Frontend: Streamlit with custom Schneider Electric branding, interactive Plotly visualizations, and responsive design
• LLM Integration: OpenAI API integration for generating natural language explanations from technical SHAP outputs
• Data Pipeline: Pandas for data processing, scikit-learn for metrics, and joblib for model persistence
• Prompting: Custom prompt templates that structure SHAP values, feature descriptions, and prediction context for optimal LLM comprehension

Challenges we ran into

• SHAP Interpretation: Translating raw SHAP values into insights that non-technical sales professionals could understand and act upon
• Performance vs. Explainability: Balancing model complexity with interpretability—more complex models might perform better but are harder to explain
• Prompt Engineering: Crafting LLM prompts that consistently generate accurate, relevant, and actionable explanations without hallucination
• Feature Engineering: Identifying which customer and opportunity characteristics would be both predictive and meaningful to sales teams
• Real-time Responsiveness: Ensuring the dashboard remains responsive when calculating SHAP values and generating visualizations for hundreds of predictions

Accomplishments that we're proud of

• Achieved F1 score > 0.70 on the test set, meeting the project's performance requirements
• Created a fully functional explainability pipeline that transforms black-box predictions into transparent, understandable insights
• Designed an intuitive interface that allows users to explore predictions at both individual and aggregate levels
• Successfully integrated LLM capabilities for on-demand natural language explanations
• Built comprehensive performance analytics with ROC curves, precision-recall analysis, and segment-based performance breakdowns
• Implemented Schneider Electric's brand identity throughout the UI with custom styling and professional visualizations

What we learned

• Explainability is as important as accuracy: Stakeholders need to understand why a model makes predictions to trust and act on them
• SHAP provides powerful insights: Feature contribution analysis helps identify which factors truly drive sales outcomes
• LLMs enhance accessibility: Combining traditional ML explainability (SHAP) with generative AI creates explanations suitable for diverse audiences
• Domain knowledge matters: Understanding sales processes and terminology was crucial for designing meaningful features and explanations
• Interactive visualization drives adoption: Sales teams engage more with tools that allow exploration rather than just displaying static results

What's next for Schneider Electric - Quick2

• Real-time Integration: Connect directly to Schneider Electric's CRM system for live opportunity scoring and explanations
• Prescriptive Recommendations: Beyond explaining predictions, provide actionable recommendations (e.g., "Adding Product A could increase win probability by 15%")
• A/B Testing Framework: Allow sales teams to test different strategies and measure impact on conversion rates
• Multi-model Ensemble: Incorporate multiple ML models and explain areas of agreement/disagreement
• Mobile Application: Deploy a mobile-friendly version for sales reps in the field
• Automated Alerting: Notify sales managers when high-value opportunities show risk signals
• Continuous Learning: Implement feedback loops where sales outcomes refine the model over time
• Extended Analytics: Add customer lifetime value prediction, churn risk, and upsell opportunity identification

Built With

• catboost
• python
• shap
• streamlit