Accelerating growth for autonomous receivables platform with a conversational AI solution

As organizations become increasingly data-driven, gaining insight from qualitative customer interactions is as critical as the quantitative side, but manual note-taking is slow, subjective, and breaks down at the scale of thousands of daily interactions across different languages and cultural cues. Collection agents generate millions of records from these conversations, making it difficult to track commitments, amounts, and other critical details from verbal exchanges.

We built a conversational intelligence solution to help collection agents make calls, take notes, and schedule follow up calls. Our engineering pod’s first challenge was to determine the best-suited tech stack to train and deliver this product on a fast-track timeline without losing accuracy. Here’s how we made it possible:

• API server: To ensure secure access to Textractive, we incorporated API key authentication as the first step in our workflow, authenticating agent IDs. Our Application Programming Interface (API) server, built using Python and Flask, performs the ID validation, authentication, and data pipeline functions.

• Define custom data pipelines: We chose Rasa, the open-source conversational AI platform, to define pipelines for entities and intents, allowing us to customize our data pipelines to meet our clients' unique needs. Rasa's plug-in-based architecture seamlessly integrated with our overall product architecture, making it our top choice. Additionally, Rasa's flexibility allowed us to swap out ML models as needed, from BERT to Open AI GPT-3.

• Training the ML model: Zemoso analyzed some client-customer conversations to identify the major categories of intent, entities involved, and the amount of the transaction. They then classified and labeled the collected samples, and generated more sample conversations for different types of intents based on the results. Using these samples, Zemoso trained a Machine Learning (ML) model.

• Intent detection: We utilized Google's Universal Sentence Encoder to encode text into high-dimensional vectors that could be classified based on intent or greetings. As a pre-trained ML model, it required less training and offered high accuracy. It enabled us to establish context by deciphering the meaning of ambiguous language in text and accelerated our go-to-market strategy.

• Categorization: We employed a Deep Neural Network (DNN) from TensorFlow for classification. TensorFlow is an end-to-end open-source library for Machine Learning (ML), widely used in the Google community.

• Entity detection: We used various techniques such as regular expressions, rules engine, and ML methods like Conditional Random Fields (CRF) that take context into account. We used Meta’s high scale Duckling library to detect and extract entities like time, ordinals, dates, numbers, and currency.

• Bringing it all together: Textractive's integration with the client’s platform has been a success. We ‌leveraged our expertise in machine learning, deep neural networks, and natural language processing to deliver a solution that provided them with a significant competitive advantage in the market.

This fintech unicorn has been able to process vast amounts of unstructured conversational data with Textractive with speed and accuracy. This kind of streamlining and amplification is significantly innovative for product companies that enable, automate, and optimize for service providers.

Zemoso combined Rasa's flexible pipeline architecture with pre-trained models for intent and entity detection. This gave collection teams a conversational AI assistant that turns unstructured, multilingual conversations into clean, actionable data, providing a meaningful competitive advantage for a fintech unicorn processing vast volumes of customer interactions.