Circiter - Data science, statistics, and AI consulting

The customer, a prominent player in the digital advertising industry in Spain, approached us with the objective of optimizing their email campaign outcomes. The project consisted of two phases: predicting customer response in terms of openings and conversions, and using these predictions to allocate emails optimally on a weekly basis, maximizing profit while adhering to various business constraints such as contact limits per customer. The goal was to minimize the number of emails sent while ensuring a certain profit level, or conversely, maximizing profit while maintaining reasonable email contact levels.

Data science

To estimate propensities, we implemented standard machine learning models with a focus on both accuracy and automatic retraining without human intervention. Dealing with class imbalance (where most emails remain unopened or receive no response) posed an additional challenge.

Allocation optimization was addressed using rarely seen, yet valuable tools in data science projects: linear programming optimization solvers. We made several adaptations to achieve reasonable execution times for large datasets with millions of variables and constraints. Through ad hoc heuristics, we significantly reduced the computing time from hours to minutes, enabling almost real-time solutions.

Technology

Leveraging the customer's existing data storage model, we developed ETL pipelines in Python (Pandas) and utilized Python's Scikit-learn for predictive modeling. We established a pipeline for periodic model retraining and employed state-of-the-art linear programming solvers to tackle the optimal allocation problem.

This recent project involved the implementation of a classic churn model with several advanced and ad hoc modifications. In addition to providing churn probabilities, we categorized the most likely reasons for high churn probability (e.g., system failures or price increases), aiding the customer success team in case management. These reasons were ranked based on actionability. Furthermore, the model was designed to predict churn well into the future, accounting for instances where churn had not yet been recorded but the customer had already made the decision not to renew the contract. This approach prioritized customer retention over achieving optimal performance on paper.

Data science

While churn models typically have straightforward data requirements and available tools, this project distinguished itself by incorporating natural language analysis of customer conversation transcripts to identify relevant clues related to churn. Customer scores were analyzed using an additional prediction explanation model to attribute them to various possible underlying causes, thereby facilitating subsequent retention efforts.

Technology

The entire project was implemented in R, covering data extraction, modeling, and the refitting pipeline. Also, a web application was developed to launch on demand trainings, integrated in GCP. This application was in charge of the model training and infrastructure scaling.

Electrical transformers at wind farms are subjected to significantly higher heat-related material fatigue compared to standard transformers. This is attributed to their load variability, resulting in frequent cycles of heating/cooling and subsequent expansion/contraction due to wind variability. Our customer, an international company operating wind farms across multiple countries, approached us with the task of nowcasting transformer operating temperature. This involved analyzing various climatic and operational variables to detect deviations and enable proactive maintenance actions.

Data science

The fundamental models used in the nowcast system followed a standard approach: based on the current state of the system and relevant past data (within the last 2-3 hours), the objective was to predict the present operational temperature. The primary focus lay in identifying a concise set of variables that effectively captured the system's status and considering the historical duration that influenced the current state. After selecting these variables, a classic and straightforward regression model was found to meet the operational requirements set by the customer. Furthermore, through meticulous analysis of model failures, it was revealed that incomplete climatic data and the omission of specific events (such as episodes of rain) had an impact on the accuracy of temperature prediction, leading to a reevaluation of the data gathering process.

Technology

In response to the customer's requirements, we successfully implemented a full end-to-end pipeline using Python, covering data acquisition, cleaning, and prediction. This tailored solution met their specific needs, ensuring a seamless flow of data and efficient handling throughout the entire system.

The customer was a consulting company specializing in assisting sellers across various online markets, particularly on platforms like Amazon. They sought an optimized bidding system to enhance their online performance.

Data science

Bid optimization is a relatively new field, often utilizing Bayesian models to handle data uncertainty, especially for new products or advertisements. However, the computationally intensive nature of fully Bayesian approaches necessitated a fast approximation using INLA (Integrated Nested Laplace Approximation). By quantifying estimate uncertainty, we divided the ad lifetime into an exploration phase, investing money to evaluate ad performance, and an exploitation phase, implementing an optimal policy to maximize profits.

Technology

The customer provided us with a well-designed MySQL database, allowing easy querying via Python, while modeling and analysis were performed using R and INLA. The solution was efficiently dockerized and deployed on the customer's servers.

A prominent player in the retail industry sought to analyze customer behavior within their physical stores. They had deployed azimuthal cameras capable of detecting activity, but the camera outputs were raw numeric strings associated with pixels. Our task was to transform this data by: 1. mapping camera pixels to specific points on the store layout; 2) aggregating information from multiple cameras capturing the same store area, and 3) generating visualizations, such as heatmaps, to depict store activity patterns over time. By fulfilling these requirements, we provided the client with valuable insights into customer behavior, enhancing their understanding of store dynamics.

Data science

The initial setup process for a new store was time-consuming. However, we significantly expedited it by leveraging a model of the camera optics and image deformation, specifically designed for bull's-eye cameras. Although limited inputs were available (requiring some human intervention), we employed standard nonlinear models capable of learning the camera orientation and deformation parameters. This approach allowed for more efficient and accurate setup procedures, significantly reducing the time and effort required for each new store.

Technology

The solution we developed found successful implementation in two major retailers in Spain. To support this, we deployed the solution on Google Cloud, utilizing Compute Engine, Cloud Functions, and BigQuery Users were able to manage the store information using a Python Flask web application, and access hourly statistics, maps, and heatmaps through a high-performance dashboard built with R Shiny technologies. This ensured efficient and seamless access to valuable insights for decision-making and monitoring purposes.