Break Through Tech AI Studio | Host Company: Salesforce

Smart CRM Helper is an intelligent conversational AI assistant built for Salesforce sales teams to reduce manual CRM work, surface high-value opportunities, and enable data-driven decision making through predictive analytics and natural language interaction.

• Built three supervised machine learning models for lead scoring, opportunity win prediction, and account health scoring using Random Forest, Gradient Boosting, and ElasticNet.
• Achieved strong performance across tasks, including approximately 94 percent accuracy for lead scoring, approximately 96 percent accuracy and a 97.3 percent F1 score for opportunity win prediction, and an R² of 0.9466 for account health scoring.
• Designed a conversational AI workflow using sentence transformers, ChromaDB, Ollama, and Gradio to enable natural language CRM queries.
• Delivered measurable business value by reducing manual data search time and improving sales prioritization and retention outcomes.

Since this is a private repository, you will clone it directly to your computer.

1. Open your terminal or command prompt.

2. Navigate to the folder where you want to store your project.

3. Clone the repository using the following command. Make sure to use the correct URL for your project.

   git clone [https://github.com/ishween/Team2A.git](https://github.com/ishween/Team2A.git)
   
   

4. Change into the project directory you just cloned.

   cd Team2A
   

Python dependencies:

pip install pandas numpy matplotlib seaborn scikit-learn xgboost sentence-transformers chromadb jupyter notebook ipykernel gradio

Ollama installation

1. Download Ollama from: https://ollama.com/download
2. Pull model to be used:

ollama pull llama3

Datasets are inlcuded in the repo

data/processed

Start Jupyter Notebook

jupyter notebook

Execute the notebooks in order: account health, lead scoring, opportunity win, sentence transformer, and then agents.ipynb inside the agents folder

This project was completed as part of the Break Through Tech AI Studio, a workforce development program that partners students with industry companies to solve real business problems using machine learning and AI.

Host Company: Salesforce
Industry: Customer Relationship Management (CRM)

Sales teams spend a significant portion of their time manually searching CRM systems, prioritizing leads, and identifying at-risk accounts. This leads to lost opportunities, inefficient workflows, and reactive decision making.

Smart CRM Helper transforms Salesforce from a passive data repository into an active decision-support system by:

• Automatically scoring and ranking leads
• Predicting opportunity win probabilities
• Identifying at-risk and high-value accounts
• Allowing users to ask natural language questions such as:
  Which healthcare accounts are likely to close this quarter?

https://www.kaggle.com/datasets/innocentmfa/crm-sales-opportunities

• Source: Salesforce CRM data
• Size: Approximately 8,800 sales opportunities
• Structure: Four merged tables including accounts, products, sales pipeline, and sales teams

Key preprocessing steps included:

• Handling missing values and duplicate records
• Feature engineering for engagement, recency, and performance metrics
• Standardization and encoding of categorical variables
• Correlation analysis to prevent multicollinearity and data leakage

• Revenue, recency, and engagement metrics showed the strongest correlation with account outcomes
• Severe class imbalance required careful metric selection and validation strategies
• Time-based splits were necessary to ensure temporal generalization
• Distribution of closed vs. non-closed deals highlights significant class imbalance, motivating the use of F1 score and recall-focused evaluation.

• Revenue scales non-linearly with company size, motivating log transformations for revenue and employee count.

• Strong correlations between revenue and employee-related features informed feature selection and regularization strategies.

• Pre-model feature correlations guided feature selection for the opportunity win prediction model.

Annotated visualizations, including feature importance plots and confusion matrices, are included within the notebooks.

• Lead Scoring: Random Forest was selected for its ability to capture non-linear interactions while maintaining strong precision.
• Opportunity Win Prediction: Gradient Boosting provided the best balance of accuracy, recall, and robustness.
• Account Health Scoring: ElasticNet handled correlated features and enabled interpretable weighting of business drivers.

• Stratified cross-validation for classification tasks
• 80/20 train-test splits
• Time-based validation for regression modeling
• Metrics included Accuracy, Precision, Recall, F1 Score, ROC AUC, and R²

Model	Key Metrics
Lead Scoring	Approximately 94 percent accuracy
Opportunity Win Prediction	Approximately 96 percent accuracy, 97.3 percent F1 score, 99.7 percent of wins identified
Account Health Scoring	R² of 0.9466, time-based R² of 0.9324

These results exceeded baseline expectations and demonstrated readiness for real-world deployment.

• lead_scoring.ipynb: Feature engineering and Random Forest training pipeline
• opportunity_win.ipynb: Gradient Boosting model with leakage prevention
• account_health.ipynb: Synthetic target construction and ElasticNet regression
• sentence_transformer.ipynb: Embedding generation and ChromaDB vector storage
• agents.ipynb: Agentic routing and Gradio conversational interface

What worked well:

• Clear separation of predictive tasks
• Strong alignment between business objectives and model outputs
• Effective integration of NLP with traditional machine learning

Challenges:

• Lack of ground truth labels for account health required synthetic target design
• Class imbalance demanded careful evaluation strategy
• Deployment constraints limited real-time integration

• Add richer explanatory visualizations and outputs
• Develop an interactive dashboard or web application

This project was completed for educational purposes as part of Break Through Tech AI Studio. No open-source license is currently applied.

We thank Ishween Kaur, Challenge Advisor, and Leah Dsouza, AI Studio Coach, for their guidance and support throughout this project.