Sales Conversion Model Reinf Learning

🚀 Sales Conversion Prediction Model

This is a reinforcement learning model trained to predict real - time sales conversion probabilities during customer conversations using Azure OpenAI embeddings, which provides valuable insights for sales processes.

🚀 Quick Start

Run On Colab With unsloth/gemma - 3 - 4b - it - GGUF model and BAAI/bge - m3 opensource embeddings

✨ Features

• Trained for real - time sales conversion prediction during customer conversations.
• Utilizes Azure OpenAI embeddings for state representation.
• Can identify key conversation patterns for different types of customers.

📦 Installation

⚠️ Important Note

It's crucial to use Python 3.11 version for inference; otherwise, it will cause layer dim errors.

conda create -n salespy11 pip python=3.11
conda activate salespy11
pip install azure-openai stable-baselines3 numpy torch huggingface_hub
git lfs install
git clone https://huggingface.co/DeepMostInnovations/sales-conversion-model-reinf-learning
cd sales-conversion-model-reinf-learning

💻 Usage Examples

Run Open - Source Model Directly

!python opensource_inference.py --model_path sales_conversion_model

Run azure embedding trained model

⚠️ Important Note

Use Python 3.10 version for inference; otherwise, it will cause layer dim errors.

conda create -n salespy10 pip python=3.10
conda activate salespy110
pip install azure-openai stable-baselines3 numpy torch huggingface_hub
python sales_inference.py --model_path sales_model.zip --conversation_json sample_conv_generated.json --azure_api_key "api_key" --azure_endpoint "https://resource_name.openai.azure.com/" --azure_deployment_name "text-embedding-3-large" --azure_api_version "2023-12-01-preview" --embedding_dim 3072

Output

--- Conversation Predictions (with Azure OpenAI Embeddings) ---
Turn 1 (customer): "Hello, I'm looking for information on your new AI-powered CR..." -> Probability: 0.1744
Turn 2 (sales_rep): "You've come to the right place! Our new AI CRM is designed t..." -> Probability: 0.3292
Turn 3 (customer): "maybe not for us. its a great product, in future may buy..." -> Probability: 0.3350
Turn 4 (sales_rep): "Excellent, those are two key strengths. Our AI analyzes lead..." -> Probability: 0.3908
Turn 5 (customer): "looks oke, but maybe we can't consider..." -> Probability: 0.3833

Generate And Clean Dataset

python generate_dataset.py --num_conversations 100000 --num_profiles 20 --output_path custom_dataset.csv --num_threads 15 --rate_limit 2000 --batch_size 10
python clean_dataset.py custom_dataset.csv --chunk_size 1000  --skip_encoding_check

Training

git lfs install
git clone https://huggingface.co/datasets/DeepMostInnovations/saas-sales-conversations
cd saas-sales-conversations
python train.py --dataset cleaned_custom_dataset.csv --model_path {sales-conversion-model-reinf-learning path}/sales_model --timesteps 200000  --batch_size 64 

📚 Documentation

Model Architecture

Training Data

• Synthetic sales conversations generated using large language models.
• 100,000+ conversation scenarios across different customer types.
• Embeddings captured conversation semantic meaning.

Model Performance

The model learned to identify key conversation patterns:

• Technical buyers respond to detailed features.
• Price - conscious customers need ROI justification.
• Early - stage prospects require needs assessment. According to the paper, SalesRLAgent achieves:
• 96.7% accuracy in conversion prediction.
• Outperforms LLM - only approaches by 34.7%.
• 85ms vs 3450ms inference speed compared to GPT - 4.
• 43.2% increase in conversion rates when used by sales representatives.

Metrics

📄 License

MIT License - Feel free to use and modify for your needs.

📚 Citation

If you use this model, please cite:

@article{nandakishor2025salesrlagent,
  title={SalesRLAgent: A Reinforcement Learning Approach for Real-Time Sales Conversion Prediction and Optimization},
  author={Nandakishor, M},
  journal={arXiv preprint arXiv:2503.23303},
  year={2025},
  url={https://arxiv.org/abs/2503.23303}
}