Case Study: Empowering Users of a Building Materials Trading Platform with a Smart, AI-Powered Assistant

Disclaimer: Please note that this case study on AI product is presented in an anonymous format due to a non-disclosure agreement (NDA) with our client.

The Client

The client, a European startup active in the bulk shipping sector with operations across the Mediterranean and African regions, benefits from the expertise and backing of experienced stakeholders, establishing a solid foundation for its business activities.

The Product

The trading platform is a one-stop-shop, shipping operations management and materials trading platform that allows its users to buy, sell, and manage bulk materials cargo across Mediterranean shores. Developed from scratch by Boldare, it is the first digital marketplace enabling all industry players to conduct business easily, regardless of their size.

You can read the case study of the product here: Bulk shipping industry: disrupting the market with a digital platform.

One of the biggest challenges the platform’s decision-makers faced is that the industry is rather traditional:

• It operates primarily on direct business partner engagements.
• Deals are executed through phone calls and predominantly Whatsapp messaging.
• Key transaction documents typically need manual signing by concerned parties.
• Smaller enterprises struggle to access important information, affecting their operations.

The challenge

In order to make the platform more user-friendly for those users who prefer the traditional way of doing deals, our team came up with a new feature. We combined our technologically forward-thinking approach with the client’s industry knowledge to create a proof of concept for our future product — an OpenAI API-backed chatbot that allows trading platform users to make deals via WhatsApp or other interfaces.

Following our proposal to the platform’s decision-makers and subsequent discussions, we received the go-ahead to create a unique, personal, AI-powered assistant. This assistant was aimed at assisting users in navigating the platform in a way that felt familiar and comfortable, particularly through WhatsApp.

The challenge? No one in this market had developed a similar product before us.

You might be also interested in the article:

Chatbots and Virtual Assistants - Revolutionizing Customer Support

The Solution: AI assistant for business materials trading

Since we were already familiar with the building materials trading platform we had created, as well as the complex domain of industry knowledge, we could focus exclusively on developing an AI-powered product that would meet the needs of an important user segment

The journey of creating this smart assistant began in August, following a successful engagement with the platform’s stakeholders and investors. The initial phase of development was dedicated to building the app’s architecture and crafting a communication flow that would allow users to seamlessly inquire about their cargos, previously added and tracked on the web interface of the platform.

The functionality of this AI assistant was designed to encompass a variety of user interactions. For instance, it was equipped to address queries such as “Where are my cargos?“ or “When will my cargo arrive at the Tanger Med port?“ More significantly, the assistant was adept at handling complex requests, such as “I want to sell 100 tons of cement from my factory in Alexandria at US$0.15/kg, available from February 3rd.“ The ability of the assistant to interpret a question posed in multiple ways was crucial, and in cases where the user’s query lacked certain details, it was programmed to solicit the missing information.

Additionally, it played a vital role in streamlining the processing of these requests through the platform, thereby enabling users to efficiently complete their transactions. The assistant also had the capability to formulate commercial proposals for other users via straightforward dialogues.

Our primary goals for this AI assistant for business materials trading were:

• To ensure that the smart assistant had constant and secure online access to the platform and its user data, thereby enabling it to perform actions on behalf of the users.
• To empower users of the platform to conduct commercial transactions and engage in conversations on relevant industry topics with the assistant.
• To provide users with convenient access to the assistant via WhatsApp chat.
• To ensure that the language skills of the assistant were versatile enough to perfectly comprehend users and their intentions, catering even to those for whom English is not a first language.
• Despite the challenges of having limited information on the operational aspects of the app and no pre-existing functional benchmarks, our team succeeded in developing the full functionality of the smart assistant using our innovative approach and unique architecture.

This development process continued through September and culminated in a collaborative presentation with the platform’s CEO and our Head of AI Solutions at the Intercem 2023 industry conference in Istanbul.

During the development of the smart assistant, we encountered and overcame various challenges associated with GPT-4, including:

• Effective thread management.
• Efficient message history management.
• Managing the knowledge base of the platform, encompassing both writing and reading information from it.
• Recognizing user intent accurately.
• Acquiring relevant data to perform actions based on the recognized intent, such as retrieving the appropriate IDs.
• Executing the desired action, like retrieving a specific list from the API and displaying the data or fetching particular data from the API for display purposes.

The Pivot: OpenAI Assistant API

On the 6th of November, at the OpenAI DevDay, a significant revelation came to light for our project. We learned that many of the challenges we were facing, particularly those concerning the context window, were on the verge of being resolved with the introduction of a new tool — the Assistant API.

The OpenAI Assistant API is specifically designed for crafting advanced applications using OpenAI’s language models, such as ChatGPT. This tool enables natural user interactions and offers functionalities like answering questions and text generation. Our team discovered that this tool could effectively address and help us overcome many of the obstacles we had encountered up to that point.

Upon its release shortly after, we evaluated the potential of the Assistant API for integration with our Gen-AI powered assistant. The decision to incorporate it into our product marked a pivotal moment in the project. With this strategic pivot, we were able to expedite the development process and, within the following two weeks, successfully deliver the main functionalities and achieve our set product goals.

From our team’s perspective, two features of the Assistant API were particularly instrumental in the success of our smart assistant:

• The ability to handle expanded conversation contexts, providing a more seamless and natural interaction experience.
• Support for executing multiple actions simultaneously, enhancing the assistant’s efficiency and responsiveness.

However, adopting the OpenAI Assistant API wasn’t without its challenges:

• As a relatively new solution, establishing best practices for its utilization posed some difficulty.
• The API’s limitation in setting the temperature parameter, which plays a crucial role in influencing the generation of responses and can affect the quality and relevance of the output.
• Constraints regarding the number of files, the size of the knowledge base, and a lack of straightforward methods for data categorization were also concerns.

Despite these hurdles, the Assistant API proved to be the most suitable solution for enhancing the capabilities of our Gen-AI powered assistant and similar applications in development.

How does our AI assistant with OpenAI Assistant API work?

You can see the app’s logic in this diagram, which highlights the key components and their interactions. Compared to the initial version of the app we created, the Assistant API now covers most of the features that our team was managing with our internal solution.

When a user initiates a conversation, their input is directed to a dedicated Chatbot API developed by our team. This API is central to analyzing the input and generating the corresponding output. It is integrated with an Assistant API for enhanced processing capabilities.

The Assistant API employs a variety of tools to accurately interpret user inquiries, align them with the appropriate responses, and retrieve up-to-date user data from the building materials trading platform. These include algorithms for embedding and vector search to refine the accuracy of query responses and a document management feature for handling file uploads. The system also manages conversation threads to ensure a seamless and individualized communication history.

Responses are managed by a function calling handler that activates a set of established procedures in reaction to user queries. These procedures are carried out with precision, through a combination of internal actions and, if necessary, communication with external APIs to broaden the application’s range of functions.

In summary, the app works by receiving user input through WhatsApp or a user app, processing that input using the Assistant API to understand the intent and determine actions, executing those actions either within its own system or by calling external services, and, if necessary, reading from or writing to a database.