FAQ

Our platform employs a sophisticated array of predictive models to cater to a wide range of applications, from natural language processing to forecasting and beyond. Here’s a breakdown of the types of models we utilize:

1. Large Language Models (LLMs or Generative AI):
By default, our system utilizes Azure, featuring ChatGPT 4, for Large Language Models. We provide the flexibility to switch between various LLMs, including LLama2, OpenAI, Mistral, among others, to suit user preferences or specific project requirements. Our platform is designed to integrate nearly any LLM, ensuring versatility and adaptability to various tasks. If a requested LLM is not directly supported, we take steps to establish compatibility and connectivity to meet user needs.

2. Models for Classification, Regression, and Automated Solutions:
Our approach to classification and regression tasks incorporates an extensive range of models, such as Linear, Decision Tree, Random Forest, Extra Trees, LightGBM, Xgboost, CatBoost, Neural Network, and Nearest Neighbors. This diverse toolkit allows us to tackle a broad spectrum of challenges effectively.
Additionally, our system is equipped with an adaptive selection mechanism for automated solutions, dynamically choosing the most appropriate model based on the context and the desired training quality. This ensures that the models we deploy are tailored to the specific requirements of each task, ranging from simple classifications to complex predictive analytics.

For scenarios requiring heightened accuracy or performance, our platform doesn’t rely on a single model. Instead, we generate multiple models of different types and configurations, rank them based on their performance, and then create an ensemble of the best-performing models. This ensemble approach significantly enhances the accuracy of our predictions, enabling us to deliver superior results without necessitating complex inputs from users.

3. Models for Forecasting:
For forecasting problems, our toolkit includes models like Arima, Sarima, Prophet, and Bayesian models. These are selected for their proven reliability and effectiveness in predicting future trends and patterns across various domains.

Through this comprehensive and adaptive use of predictive models, our platform ensures that users have access to cutting-edge AI technologies tailored to their specific needs, simplifying the path to obtaining accurate and actionable insights.

Basically we split the user dataset (this is the source of truth) in two small sets.
Those are the Training set (generally 80% of the original dataset) and Test set (20%).

We train a machine learning algorithm in order to get good accuracy with our Training set.
What training I mean, is reducing error on predictions. In general terms initially we have a random model that basically gives you a random response, then we ask for a prediction on our training set, the model give us an answer (initially random), then we backpropagate all the needed adjustments in order to make our model answer what we expected (because we know the truth). Then we keep going with all the other rows in our training phase.
In fact, this process is done a lot of times.

After a model is trained and has a good accuracy on our training set, we just make a prediction on our testing set (unseen data for our model), then we evaluate how good is our model predicting unseen data, this is generalizing.
This testing accuracy is the one that we plot to the users (because of course the model will have a good accuracy on the training set).