Open-Source RL Model for Predicting Sales Conversion from Conversations

1 point

a year ago

For the past couple of months, I have been working on building a chess game kinda system for predicting sales conversion probabilities from sales conversations. Sales are notoriously difficult to analyse with current LLMs or SLMs, even ChatGPT, Claude, or Gemini failed to fully analyse sales conversations. How about we can guide the conversations based on predicting the conversion probabilities, that is, kinda trained on a 100000+ sales conversation with RL to predict the final probability from the embeddings. So I just used Azure OpenAI embedding(especially the text-embedding-3-large model to create a wide variety of conversations. The main goal of RL is conversion(reward=1), it will create different conversations, different pathways, most of which lead to nonconversion (0), and some lead to conversion(1), along with 3072 embedding vectors to get the nuances and semantics of the dialogues. Other fields include

* Company/product identifiers * Conversation messages (JSON) * Customer engagement & sales effectiveness scores (0-1) * Probability trajectory at each turn * Conversation style, flow pattern, and channel

Then I just trained an RL with PPO, by reducing the dimension using a linear layer and using that to do the final prediction with PPO.

Dataset, model, and training script are all open-sourced. Also written an Arxiv paper on it.

Dataset: [https://huggingface.co/datasets/DeepMostInnovations/saas-sales-conversations](https://huggingface.co/datasets/DeepMostInnovations/saas-sales-conversations)

Model, dataset creation, training, and inference: [https://huggingface.co/DeepMostInnovations/sales-conversion-model-reinf-learning](https://huggingface.co/DeepMostInnovations/sales-conversion-model-reinf-learning)

Paper: [https://arxiv.org/abs/2503.23303 ](https://arxiv.org/abs/2503.23303)

Btw, use Python version 10 for inference. Also, I am thinking of using open-source embedding models to create the embedding vectors, but it will take more time. Also I just made a platform on top of this to build agents. It's completely free, https://lexeek.deepmostai.com . You can chat with the agent at https://www.deepmostai.com/ from this website