Before we can start with a machine learning project, we must have a clear and well defined objective. What is the problem we want to solve? What is the business value? Do we have all the data that is needed to solve this problem? How is this data delivered? Should the data be anonymized? Have similar studies already been carried out in the technical literature that we can rely on? These are some of the questions that we answer in this phase.
An import prerequisite for building the model is the preprocessing of the data. We look for the best numerical (i.e. machine readable) representation for data that does not yet exist in numerical form. Then we use well-known static methods to normalize data, detect outliers and catch missing values. Furthermore, we document our findings in a report that can be used to improve the quality of your data.
The most important step in a machine learning project is building the model. We carefully select different appropriate algorithms to solve the problem as efficiently and accurately as possible. Each algorithm also has so-called hyperparameters to select, this results in an iterative process to tune them and evaluate the results. Our experiences during this process are documented in a report in which we compare and describe which methods and hyperparameters yield the best results.
Training the initial model is only half the job! Any AI model worth its salt will tell you that they need a human in the loop to provide them with periodic retraining to account for evolving data. To achieve this, you need a well thoughout data pipeline that makes sure that your AI model has access to both training and inference data.
Once the model is ready and all the data is in the right price, our brains need a jar to live in. Be it in the cloud or on your servers, we use technologies like Docker and Kubernetes to make sure our brainjars are both portable and scalable.
AI’s, just like HI’s (Human Intelligences), make mistakes!
That’s why we believe that the best AI systems should work in collaboration with humans. The AI takes care of tedious, repetitive work, while the human gets promoted to AI Manager, checking the AI’s work and correcting when necessary.
This way, we ensure that our applications continue to work outside the lab environment, and can even improve and adapt over time, ensuring that your application keeps performing for years to come!
AI is more than AI
Even though building AI models is the most fun part of our job, we believe that a good AI application is about more than just the model. With our holistic approach, we take care of everything from ideation, objective determination, backend integration, front-end design, all the way to final deployment.
This approach has made us the ideal partner for so called “Cold Start” scenarios, in which we design an AI application without any initial available data. In these scenarios, we build AI models that are completely trained by the end users. Initially, this training is quite intensive, but over time the AI model requires less and less correction until it can operate almost completely autonomously.
The right tools for the job
At Brainjar, we believe that every project requires a unique approach. Therefore, we pick the best possible solutions for every project based on the customer needs, regardless of supplier.
In this approach, we maintain the “Use when possible, build when necessary” philosophy. When possible, we use prebuilt services or products to reduce costs and development time. However, when the use case calls for a custom-built solution, we use our extensive domain knowledge and experience with frameworks like Tensorflow to build powerful custom AI models.
Let's talk about your project!
Do you want us to brainstorm about your next AI project or do you already have an idea, feel free to contact us.Contact us