How are we integrating artificial intelligence resources into the platform
Let's share with you our understanding of this new technology and how it can provide more accurate and customized insights.
The basis of everything are the great language models (LLMs)
Have you ever wondered how they boost AI tools?
The central idea is simple: these systems are trained to use enormous amounts of knowledge and “predict the next step”.
In other words, within the gigantic knowledge base of an LLM are numerous scenarios, including the most common.
Among the most common, there are the most probable, even the most probable among the probable.
That will be the result that artificial intelligence will present to you.
Vectors represent concepts such as numbers
Artificial intelligence also uses a specific type of database, where everything is transformed into a numerical code.
Instead of representing a word as a sequence of letters, for example: “cat”, LLMs transform it into a vector, which is a long list of numbers.
For example, the term “hexgrid” in Mapfry is represented by a vector that captures characteristics such as the demographic profile, location, and consumption behavior of a place.
Just as we can represent cities with geographical coordinates, the models transform words into numbers in a “vector word space”, almost like those word clouds, but multidimensional.
In this vector space, terms with similar meanings, such as “store”, “commerce” and “mall”, are close to each other.
In Geomarketing, this translates into the possibility of identifying patterns and relationships between location data and consumer behavior.
Advantages that artificial intelligence brings us
Mapfry, as a Geomarketing platform, deals with large amounts of spatial and behavioral data, which can easily be represented in multidimensional vector spaces.
A vector database specifically optimizes the representation and search for similarity in large data sets, something essential for AI-based applications, especially embedding models for recommendation or geographical clustering.
For example:
- Efficient clustering of areas for market research and business expansion
- Personalized recommendations based on similar geographic and sociodemographic characteristics
- Customer segmentation based on behavioral and demographic profiles derived from large market and transactional databases
- In identifying similar locations for business expansion, market segmentation, and recommendation based on proximity and sociodemographic characteristics
The role of transformers
The transformer is the fundamental piece that allows LLMs to understand the context.
Imagine that Mapfry receives the following information: “I am interested in clothing stores in the central region”.
The system transforms each word into vectors and, through multiple layers of transformers, “observes” the relationships between them - for example, associating “clothing store” with location data and purchase preferences.
Each layer of the transformer “refines” the understanding, adjusting the vectors so that the context becomes increasingly clear.
If the Mapfry platform is helping a shopkeeper identify where to open a new store, these mechanisms allow the AI to understand nuances, such as the difference between “center” (commercial area) and “center” (city center with residential characteristics) and thus provide accurate recommendations.
The expression “refining an AI model” (fine-tune) describes the work of qualifying a transformer to make better associations.
Knowing how to connect the dots
Inside a transformer, there are two main processes:
- Attention: Each word “seeks” other words with relevant context.
- For example, in a sentence such as “The consumer visited the store in the center and...”, the system identifies that “store” and “center” have a strong relationship, adjusting their internal representations.
- Feed-Forward: Then, each word “reflects” on the information received to help predict the next word.
- Think of it as an analysis where each element has access to the previously processed context.
In practice, if Mapfry wishes to predict consumption trends in a particular region, the AI system can use similar logic to “predict” future behavior based on historical data.
Something like identifying that, if there is a high concentration of young people in a neighborhood, the demand for specific products tends to be greater.
Learning with real data
Just as LLMs are trained with hundreds of billions of information points to adjust their billions of parameters, Mapfry's AI is being trained with data on location, demography, mobility patterns, and more.
During training, the system “adjusts” its parameters so that, starting from a prompt, such as:
Assess the feasibility of opening 10 new units in cities with more than 200,000 inhabitants.
Which neighborhoods are most suitable for opening a franchise that serves consumers with a profile similar to neighborhood X, taking into account competition and consumption patterns?
What are the areas where the average household income has been growing above the national average and which are close to shopping centers with high commercial activity?
The system is capable of:
- Generate an action plan to achieve this objective
- Perform these actions by calling additional resources
- Reflect and correct errors throughout the process
- Indicate when the task was completed and consolidate recommendations
There are different degrees of automation in the human-machine relationship, so we organized a scale inspired by the RACI Matrix, in which:
- Responsible (R): The entity that performs the task.
- Approver (A): The entity that has ultimate responsibility for the task.
- Consulted (C): The entity that provides information before execution.
- Informed (I): The entity that is kept up to date on progress or result.
Learning is real
Although the internal details of LLMs, such as attention mechanisms, feed-forward layers, and massive training, are complex, the essence is that these systems learn based on patterns found in large amounts of data.
In practice, this means that, just as the GPT can “complete” texts in a surprising way, we will use similar techniques to transform raw data into strategic Geomarketing insights.
By applying these technologies, artificial intelligence not only enhances your analyses, but also offers you the possibility of making more informed decisions, customizing campaigns and strategies according to the profile and behavior of consumers in different regions.
What exactly is Mapfry's mission