Sergio Rodríguez de Guzmán is currently the CTO of PUE, one of the leading companies in our country when it comes to implementing integration, governance and exploitation projects for business data. In recent years, the company has also embraced trends such as Generative Artificial Intelligence, where, as could not be otherwise, the quality, security and correct evaluation of data play a fundamental role.
In this interview, he talks about the most important challenges faced by organizations that want to implement this type of project, how it is advisable to carry out an audit of business data, and how the PUE itself can help companies in their digitalization processes related to a technology that is already a turning point in many areas.
(MCPRO) What are the current use cases of generative AI that can be applied across different industries?
(Sergio Rodriguez de Guzman) We are currently seeing that generative AI has potential for many sectors. A prominent example is the content generationboth text and images, although video is still in its early stages. This allows for exponential productivity increases in areas such as writing emails, product descriptions or even code generation.
For development of chatbots and virtual assistants that optimize customer service and improve internal assistance, both for employees and external customers. Another relevant application is the personalizationas generative AI can offer personalized recommendations and suggestions based on user preferences, which is useful in sectors such as e-commerce.
On the other hand, it is being used for data analysis and reportingextracting insights and detecting patterns in large volumes of data, enabling the creation of automated reports. Finally, large-scale language models help in the process optimizationallowing to automate and improve existing processes.
(MCPRO) Which sectors are most advanced in the adoption of generative AI and in which specific areas are they seeing its benefits most immediately?
(Sergio Rodriguez de Guzman)He technology sector is one of the most advanced in the adoption of generative AI. For example, code generation has seen remarkable growth, while at the same time the time required to develop unit and end-to-end tests has been drastically reduced.
Previously, such testing could account for around 30% of the total cost of a project, but with generative AI this cost has decreased significantly. In addition, the creation of user interfaces, prototypes and graphic design are accelerating considerably. On the other hand, the sector of marketing and advertising has also found great benefits, especially in the generation and personalization of content, as well as in the analysis of customer sentiment about products and brands. Finally, in the financial servicesGenerative AI is primarily used in customer service.
(MCPRO) What are the most important challenges that companies face when implementing generative AI projects?
(Sergio Rodriguez de Guzman) First of all, the Data quality and biasGenerative AI relies on high-quality data to avoid erroneous results or “hallucinations,” and this represents a great challenge due to the complexity of managing large volumes of data in current data lakes.
On the other hand, we have The infrastructure and computing resourcesGenerative AI requires a robust and scalable infrastructure, which can lead to high costs if not properly optimized. And finally, there is a shortage of specialized AI talent, which means that companies need to invest in training and acquiring personnel with specific skills to carry out these projects.
(MCPRO) What role does data quality and its proper management play in AI’s ability to generate accurate and reliable results? What are the main risks of working with poorly labelled, outdated or biased data?
(Sergio Rodriguez de Guzman)The data quality is critical for generative AI to be able to generate accurate and reliable results. If the data used in training the models is of low quality, the model will learn incorrect patterns, resulting in unreliable answers. A particularly critical problem is the bias in dataIf the dataset is biased, the results produced by the model will reflect those same biases, which can lead to serious problems, such as discrimination.
For example, a model trained solely on data from people of a particular race might have difficulty identifying people of other races. Also, when the data is mislabeled, outdated or biasedmodels learn incorrect associations, which can lead to irrelevant or inaccurate results, especially in predictive models. Ultimately, this can undermine user trust and negatively impact the perception of generative AI within organizations.
(MCPRO) What practices do you recommend companies follow to ensure their data is ready to be used in an AI project? What data validation and cleansing processes are critical in this area?
(Sergio Rodriguez de Guzman)It is critical for companies to adopt certain key practices to ensure their data is ready for AI projects. First, you need to have a clear definition of objectives of the project, which helps to identify what type of data is needed and in what format. It is also crucial collect diverse datawhich helps avoid bias and ensures that the model can generalize correctly.
He data governance is very important; data must be organized and stored in a way that is accessible and secure. As for data validation and cleansing processes, it is important to ensure that there are no duplicates, correct typographical errors and atypical data, and ensure that the information is true and complete.
(MCPRO) How does a well-defined data strategy help reduce “hallucinations” in generative AI models? What types of hallucinations are most common when data management is inadequate?
(Sergio Rodriguez de Guzman)A well-defined data strategy is key to reducing hallucinations in generative AI models. This involves having high-quality data, proper cleaning and pre-processing processes, accurate labeling, data diversity, and constant updates.
The most common hallucinations often include nonsensical responses, incorrect information, biases, factual errors, and logical inconsistenciesThese errors occur when the model faces limitations in the input data or in its pre-training, resulting in responses that may appear credible but are actually incorrect.
(MCPRO) What mechanisms or audits can be implemented to ensure that the data used to train AI remains accurate and relevant over time?
(Sergio Rodriguez de Guzman)To ensure that data remains true and relevant, it is crucial to implement a continuous monitoring. This can include dashboards and alerts that allow real-time visualization of data quality, as well as detecting any significant deviations. Another useful technique is the cross validation with independent test data, which can help identify any deterioration in model accuracy.
Not everything is technology, also the audits performed by people that randomly review data quality are essential to maintaining adequate standards. And the user feedback It is very valuable, as it allows reporting possible errors or incorrect information generated by the model.
(MCPRO) What security and privacy measures should companies consider before starting a generative AI project?
(Sergio Rodriguez de Guzman)It is important to make a comprehensive risk assessmentidentifying the sensitive data to be used and potential vulnerabilities. It is also necessary to ensure that the selected providers meet high standards of security and privacy. During the development of the project, measures such as data encryptionaccess control and continuous monitoring.
It must be ensured that only strictly necessary data is collected and stored, using anonymization and pseudonymization to protect the identity of users. Complying with current regulations and being transparent in the use of data is essential to ensure a secure environment.
(MCPRO) Finally, what kind of support does PUE Data offer to companies that want to launch a generative AI initiative?
(Sergio Rodriguez de Guzman)At PUE Data, with over 10 years of experience in large-scale data projects, we have made generative AI the centerpiece of our service offering. The success of generative AI and projects based on it depends on the design, construction, optimization and advanced management of data lakes, an area in which PUE Data is a reference at EMEA level. In addition, we have experience in real use cases, which allows us to guide companies through a dynamic and adaptable process, thus ensuring that their generative AI projects are successfully developed in an ever-changing technological environment.