What is an AI prototype?

Prototype and Proof of Concept (PoC) have been established concepts in both physical and digital development for many decades. Now, these terms are also gaining traction in the AI domain, and it’s becoming increasingly common to "prototype" models to evaluate technical options before deciding to invest in a full-scale AI project.

In this article, we delve into AI prototypes, explaining how they differ from traditional digital prototypes you might be familiar with, and present examples of some of the prototypes we have developed in collaboration with our clients and partners.

‍

‍

What is an AI prototype?

As you may already know, a Proof of Concept (PoC) is generally defined as an early iteration of a solution, for instance, a digital product, designed to demonstrate the viability of an idea or concept. A prototype typically encompasses additional steps and criteria, such as the ability to be tested, availability to a broader audience, or the capability for analysis, thereby moving the system closer to a complete solution. In the case of an AI prototype, it specifically refers to a technical prototype in which one or more components of the implementation are, or are integrated with, an AI model.

"An AI prototype can either be a standalone AI model with a specific task or a prototype of a typical software product, with a backend and a frontend, where part of the implementation involves an API integration with an AI model," says Simon Karlsson, partner at Violet who has developed the architecture for several AI solutions.

Prototypes are valuable in numerous AI projects. For instance, you can prototype a matching engine that leverages a large language model (LLM) to connect job seekers with relevant job ads, or image and video recognition models that utilize greenhouse cameras to offer optimization suggestions to boost harvests.

‍

‍

Another example of an AI prototype is a forecasting model designed to predict demand for a particular service or product through machine learning. Such a forecasting solution prototype could be quite basic, comprising solely a machine learning model that has been trained on a limited set of data, like a singular time series.

Unlike typical software prototypes, the focus is on demonstrating and testing AI-specific features and capabilities, i.e., how the model performs, while a standard digital prototype emphasizes other aspects of software and product design, such as user interface or specific functionality.

‍

What is the Purpose of an AI Prototype?

There are various reasons for creating an AI prototype. Often, it's about testing the technical feasibility of the model or solution, for example, "Can GPT-4 be used as a reasoning engine in our AI agent to respond to customer support inquiries?". The process also includes evaluating different technical options, such as "How does GPT-4 perform compared to LLaMa 2 as a reasoning engine in our AI agent?".

For large language models like GPT/OpenAI, Gemini/Google, and Mixtral/OpenSource, setup and server costs, as well as costs for usage, play crucial roles. Data security also becomes a critical consideration in deciding whether a third-party solution is viable, influencing the overall costs. In prototype development, evaluating the use case and performance against the business case for various model alternatives is a common practice.

An AI prototype can also serve communicative and organizational purposes. By using a prototype, feedback can be gathered from future users to enhance the solution. A prototype can also act as a tool to build a business case, convincing decision-makers of the system's value, which can help make a well-informed decision about investing in a full-scale production project.

Although an AI prototype is far from a finished, integrated solution, it can still provide valuable insights into what is practically and technically feasible, economically viable, and aligns with the organization's needs. The cost of a prototype could be as little as 5-10% of a finished solution, and sometimes even less. The investments made in the prototype also contribute to the final production project, both through the developed software code and through the experiences and knowledge gained during the prototype phase.

‍

Examples of AI Prototypes

1. Large Language Model as a Reasoning Engine in an AI agent

A company within nutritional supplements wanted to develop an intelligent agent capable of acting as a customer support agent and a domain expert (an AI agent). The company operates an e-commerce platform with a vast array of products for various purposes, they also have a complex guide with lots of product information. The company's customers were seeking personalized product recommendations, a service that would be costly and difficult to scale with human support agents.

The company wanted to explore the possibility of using large language models, together with their internal document library and the latest research in the field, to develop an AI agent that their customers could interact with.

‍

‍

"A critical technical aspect during the prototype phase was to minimize what are known as hallucinations, instances where the model generates inaccurate information. Therefore, ensuring the agent's responses were grounded in the specific data set was paramount," says Max Sonebäck, model developer at Violet.

The prototype was built on the company's internal data with an API integration to OpenAI and their GPT models. Language models played a key role in the system, steering the AI agent's decision-making process in what data to use and in generating summaries and compilations. Another important dimension in the prototyping phase was costs for using external AI services (OpenAI/GPT models). The client wanted an understanding and predictions of the system's future costs as the service expands to hundreds of thousands of users.

A basic cloud environment was built to enable the company and selected users to test the system. By testing the prototype, the company was able to decide on a larger investment in a full-scale AI agent.

2. AI Forecasts Predict Fuel Demand

Another client, operating in the fuel sector, believed that an AI tool capable of forecasting product demand could enhance their logistics efficiency and optimize capital allocation.

‍

‍

To estimate fuel demand for a given time period, the company previously relied on basic historical data and estimates from experienced employees - a manual approach with a high margin for error.

With the prototype, they wanted to explore whether an AI solution could provide accurate demand forecasts by considering patterns in historical data, including weekdays, months, and holidays.

"A crucial aspect of the prototyping phase was also to test and evaluate various forecasting methods, such as statistical, programmatic, and AI-based solutions. A simple problem rarely demands a complex AI solution, but in this case, an AI solution was exactly what was needed", says Karin Johansson, Machine Learning Engineer and forecasting specialist at Violet.

After confirming that the AI solution provided accurate forecasts, the decision was made to develop a comprehensive forecasting tool with an integrated AI-driven forecasting model.

‍
3. AI-driven Friend Matching and Customer Segmentation

One of the largest social networks for women in the Nordic region wanted to explore how AI solutions could enhance the user experience and increase revenue streams. The app, which matches women for friendships, had a vast amount of user-generated text data that they were not analyzing or leveraging.

"As the app's user base expanded to more than 200,000 individuals, the importance of comprehending the customers through data became more significant. Similarly, the need to leverage these insights in conversations with advertisers grew," states Anna Rydin, Machine Learning Engineer and NLP specialist at Violet.

The client saw an opportunity to uncover patterns in user data by better segmenting the customer base, allowing for more targeted content, events, and marketing from the platform's partners. Moreover, data from profile descriptions, group descriptions, and activity descriptions could be utilized for more intelligent user matching.

‍

‍

During the prototype phase, a text classification model was developed to segment the user base based on text data from profiles, activities, and groups. Language models, trained on the app's data, were used to extract and classify keywords and associate them with similar words in the texts. The prototype phase also involved comparing various technical solutions, with the final solution incorporating large language models from the research unit within Kungliga Biblioteket, KB-labb.

After the prototyping-phase, the decision was made to invest in a comprehensive, integrated AI solution for text classification and segmentation of the customer base.

‍

An Essential Component of an AI Project

As the possibilities with AI grow exponentially, there's also a need for efficient tools to easily test and validate the solutions. Prototyping AI models will continue to be crucial for exploring different technical options, ultimately leading to powerful and value-creating AI solutions.

An AI prototype also helps identify the use cases within the business where AI can add the most value in the shortest time. Therefore, prototyping is a fast and productive way to kick start the company's AI journey.
‍

‍

‍

Do you and your team need help in developing an AI PoC or AI Prototype? Contact me at mattias.guilotte@violet.ai.

‍