Almost one year ago, we launched our brand new AI platform, SAIBRE. In this post I explain how SAIBRE stands out from the competition by providing a complete AI ecosystem.
Creating any new platform requires difficult decisions to be made. When I was tasked with creating a new AI platform from the ground up, I had to choose between three paths. The first path was to try and outperform all our competitors in feature comparison tables. The second was to try and deliver the most powerful AI platform on the market. But I went with the third option, namely, to create the most flexible and easy-to-use AI platform out there. I chose the latter option, and the result is the Sonasoft AI Bot Runtime Ecosystem, or SAIBRE.
The AI ecosystem
AI platforms essentially exist in order to help create and deliver AI models. A pure AI platform exists as part of a larger AI ecosystem as shown below.
Each element in this ecosystem has a particular role to play. Let’s explore them in turn.
Data sources
AI can only exist because of data. An AI solution ingests data, processes it, and delivers an output. In each case, you will need to leverage different data sources. Among others, these include databases, sensors, video feeds, or external data APIs. For instance, to train a classifier to recognize images of cats, you will need millions of photos of cats and other animals. And to train an algorithm to forecast stock levels, you will need all your historical sales data along with data that might affect sales like historical weather data and fashion trends.
Data platform
In most companies, the data is spread across multiple systems in different formats. Some may be in legacy systems, other data may come from IoT sensors or external data feeds. All this data needs to be imported into one data platform. This is a specialized cloud system that can handle the considerable volume of data you will need. Typical data platforms also offer tools to help you clean up the data and transform it ready for use by the AI platform. This is known as data engineering or feature engineering and is a key step in the process.
AI platform
The core element of the system is the AI platform. This performs three tasks. Firstly, it allows you to take the data feed from the data platform and use it to train machine learning (ML) models that typically fall into one of four categories:
- Classification: This photo is a cat chasing a mouse
- Forecasting: This year, sunscreen sales will peak in the last week of June
- Anomaly detection: Your should see a doctor because your heart rate is irregular
- Knowledge discovery: These patents are similar to the one you are writing
Secondly, it offers tools to allow you to validate the performance of these models. Potentially, this may include the ability to do model competition to find the best-performing model. Thirdly, it should offer tools to simplify the deployment of the newly-trained ML model.
Deployed solution
By itself, an ML model is useless. You need to deploy it into your systems in order to be able to leverage it. There are various ways to do this. You might run it within your backend system and communicate with it via APIs. You may choose to install it within edge devices (as long as they are able to run the model). Or you might actually run it from the AI platform and connect via APIs. For simplicity, we refer to these deployed AI solutions as AI bots. This is often where AI transformation projects come unstuck—people simply fail to deliver a successful deployment.
Monitoring
The final critical part of any AI solution is monitoring. Any system you rely on should be monitored for reliability and performance. But with AI this is doubly important. That’s because ML models drift over time. This results in their accuracy declining. Two factors come into play. First, the external environment is evolving, meaning the data originally used to train the model has changed. For instance, the economic climate may have taken a sharp downturn. Second, most businesses undergo a degree of internal transformation simply by deploying a successful AI system. This means efficiency and productivity increase, and in turn, the ML model’s inputs change. The upshot is, you need to be able to monitor how well the model is currently working.
The four pillars of SAIBRE
Starting from a clean sheet of paper allowed me to design SAIBRE to provide most of these requirements in a single system. The only thing SAIBRE doesn’t include is your data sources! And even then, its data platform makes it very simple to connect different sources to the system. The following are the key parts of SAIBRE.
Dataset pipeline
Our data platform is designed to provide zero-effort data engineering. That means it lets you perform all the typical data science and data engineering tasks you need prior to creating an ML model. This is achieved by chaining together simple building blocks like data import, data translation, deduplication, and feature engineering. This approach allows you to create rich data pipelines that generate reusable AI datasets.
AI platform
The core of SAIBRE is the AI platform. This provides zero-effort model creation using the same drag-and-drop approach as the dataset pipeline. You chain simple modules together to create a complete AI pipeline from dataset to validated model. You can even train multiple models in parallel to allow you to choose the best performing one. For more advanced users, the system allows you to drop in custom AI model code. This makes it extremely flexible and powerful.
Deployed AI bots
The name SAIBRE actually means Sonasoft AI Bot Runtime Engine. As you can guess, that means it is designed to run your AI bot as well as create it. This ushers in the era of zero-effort deployment. You can literally deploy the bot with a single click. The system allows you to swap between development and production datasets. It also includes a lightweight version optimized for running in edge devices, such as smart sensors and the like.
Smart monitoring
The final element in SAIBRE is our unique smart monitoring system. This really sets us apart from competitors. It uses its own AI to constantly validate the health of your running AI bots. This includes checking whether the accuracy is starting to drift too far. If so, it will flag this up and let you retrain the model on the fly. It also checks your data feeds to identify potential problems. It alerts you if it sees that a data feed has gone offline or become corrupted. In short, this means you get zero-effort maintenance for your AI bot. Currently, we are working to build in smart security monitoring to check for potential hacks or attacks on your AI solutions.
If you want to learn more about how SAIBRE works, download our white paper today.