Botmeni: Q&A system for e-commerce with AI

Botmeni was created before the boom of LLMs, so our team had to learn advanced Natural Language Processing (NLP) techniques to solve the problem. To achieve this, we analyzed thousands of questions from different Mercado Libre stores, labeling them to identify the intent behind each one—whether the inquiry was about stock availability, shipping costs, business hours, etc. This initial analysis and labeling were crucial, as they helped us train various machine learning-based NLP models. Additionally, data was labeled to extract other types of information, such as user-provided locations in questions or certain product description attributes like dimensions. Both tasks were performed using Named Entity Recognition (NER) techniques. During this research phase, we evaluated multiple models, ranging from CNNs for text to more complex and capable architectures like LSTM, Bi-LSTM Attention, Bi-LSTM CRF, and BERT. As a result, we developed multiple models capable of extracting information and intent from questions, laying the foundation for further development. With the core problem solved, we proceeded with building the full platform, integrating it with Mercado Libre, adding a payment gateway, and refining the response generation algorithm. When the LLM boom arrived, we leveraged our prior knowledge to create a fusion between our existing models and LLMs, achieving the best of both worlds. Today, Botmeni features various models trained with different technologies, as well as integrations with multiple services.