Konfío

At Konfío, our mission is to empower Mexican entrepreneurs through an ecosystem of products designed exclusively for the growth of their businesses. Achieving this goal requires a vast amount of data and a highly qualified team capable of manipulating this data and translating it into valuable insights. This is why we seek talents with diverse skills and varied academic backgrounds, as well as a robust infrastructure that supports the entire process of discovery, interaction, training, and deployment of models. From the beginning, we have utilized cloud technology services. In my experience, AWS has been the most used; it fits well with our way of working and has been very flexible in solving customer problems using technology.

The philosophy at Konfío has always been and will continue to be to solve problems using the most suitable technology for each solution. This approach has allowed us to grow and evolve, especially in the area of data science.

First Steps

We started with an EC2 instance that was activated with a cronjob, downloaded transactional data from clients and all the additional information needed, and executed a Python script that evaluated fraud policies. In parallel, credit policies were evaluated by an analyst. The final decision was made using both. Soon, the need to automate the manual part of the work became evident to scale the operation, and we achieved this with Lambda functions (serverless), all in one script. We also implemented a CI/CD process to adapt policies according to payment behavior, results, and patterns we identified in the data. At that time, we used Jenkins and Gitlab, already configured and used by other teams.

Scaling the Solution

We trained various models leveraging the multiple data sources we already had. We continued to implement them with lambda functions for their practicality and low cost. But this time, we modularized the solution, separating data acquisition from feature generation and inferences. For initial orchestration, we used step functions but evolved towards our own system as needs became more complex. The demand for models that operated in real-time increased significantly, along with the need to train models in batches. In the absence of Sagemaker pipelines at the time, we opted for solutions from other providers, adapting to the tools available while our company grew and the team expanded to continue making a significant impact.

The volume of data grew even more, making local model training impractical. The transition to the cloud then became an urgent need, allowing us to perform exploratory data analysis (EDA) and training more efficiently. This migration enabled us to create specific domains adapted to the different uses our data scientists needed, offering the flexibility to choose instances according to the required characteristics for each task. Having our own cloud account allowed us to manage our costs more effectively, monitoring spending and evaluating the real impact of the solutions implemented.

 The most important thing is to focus on the customers, the problems to be solved, and the feasibility of the solutions, always counting on the talent capable of executing ideas to solve these problems  

A key piece in accelerating model training and reducing deployment times was the creation of a feature repository. This repository facilitated the reuse of features among different models and teams, significantly speeding up the development process. Additionally, the formation of a dedicated engineering team relied on this repository not only to shorten deployment times but also to strengthen the continuous integration and delivery (CI/CD) pipelines, with unit, performance, and integration tests required to minimize errors. These pipelines not only accelerated the deployment of models but also ensured the quality and reliability of the data needed for inferences.

The ability to quickly enable new solutions and correct errors efficiently was greatly enhanced by cloud infrastructure, leveraging services like AWS SageMaker (for feature storage and model registry), serverless functions such as Lambda and API Gateway, and development tools like GitLab. The implementation of these technologies and methodologies allowed Konfío not only to efficiently scale its operations but also to stay at the forefront of implementing data science and artificial intelligence solutions.

Artificial Intelligence Era

By the time the wave of Generative AI hit the market, Konfío had already made strides in this field. Various projects have been developed for sentiment and emotion analysis using transcriptions and audio tones, as well as image recognition to contain fraud and the automation of key information extraction from documents to facilitate customer experience. But with this new technology and the exploration of its different variants to identify the most suitable for our problems, we have accelerated their resolution. We used AWS BedRock, as well as OpenAI and other options in content and image generation projects (customized for our clients, but with Konfío's tone) for our Marketing teams, as well as to automate and speed up the audits carried out by operational areas.

Conclusion

We have advanced from educated guesses to more complex patterns, from training models with little data locally to requiring cloud instances with GPUs to process data. We do not limit ourselves to a single technology; we are always experimenting. AWS has been useful, but we also recognize the strengths of GCP. The most important thing is to focus on the customers, the problems to be solved, and the feasibility of the solutions, always counting on the talent capable of executing ideas to solve these problems. The technology we use comes afterward. This has been our philosophy for the first 10 years and will continue to be so in the future.