Data Science

We leverage cutting-edge techniques, from optimization to machine learning, to ensure that your business is driving value through advanced analytics for years to come.

Our approach to Data Science

$82M

Generated $82M + in value

17%

Average 17% improvement in operations efficiency

A proven process to solving Data Science problems

Client Engagement

A crucial step in the modeling process is engaging with our clients to ensure rigorous definitions around business needs and asks. Often, solutions are developed that solve the wrong problem. A solid interface and documentation ensure that the solution blueprint is executed to exactly meet client needs, and allows our team to map out what information can be included and what types of technical solutions may be appropriate.

Exploratory Analysis

After understanding business needs, the next task is to thoroughly understand the data available. This includes discussing outlier strategies, understanding relationships between variables, and feature engineering. Taking an in-depth look at the data allows us to understand what types of models and modeling flows can be utilized, and how to provide both high explainability and performance.

Model Development

Once the data has been cleaned and prepped for algorithmic solutioning, the scientific method is utilized to design experiments to understand how to best solve the problem at hand. Certain problem types require certain experimental construction. Whether it be selecting your validation set to span several seasons or cross-validating an imbalanced dataset, our team brings industry experience to the table to ensure the scientific soundness of all deliverables.

Deployment and Automation

A common oversight in data science solutioning is neglecting this piece of the equation in favor of algorithmic complexity. The only way statistics can assist clients is by ensuring smooth and valuable integration with business processes.

Case studies

Data Science

Steel industry

Projected 25% increase in throughput on the mill floor through Simulation and Optimization

Client Challenge

The client had many steel coordinators on the mill floor making decisions based on heuristic knowledge and spreadsheet-based modeling. Operators leveraged simple averages to understand possible dwell times and transit times, often overlooking large amounts of available details. The overarching goal was to keep the casting units running constantly and at maximum speed. To do this the plant floor operators needed optimal production schedules to be produced to maximize output.

Our Approach

The Aimpoint Digital team worked within the client Hadoop environment, leveraging Python, Spark, and Pyspark to take advantage of edge computing nodes to implement a Discrete Event Simulation which scaled to compute thousands of probability densities efficiently and quickly. This allowed the team to run a “what-if” style analysis to determine optimal schedules given production requirements. The Aimpoint team was also able to include additional functionality to the client team, layering in scheduled machine downtime; the developed solution consists of the option for a JSON push detailing the specific downtimes to understand the effect of machine maintenance on optimal production schedules.

Outcome

Reduced total time to complete production schedules an average of 25%, meaning a 25% increase in plant productivity. Client team is now able to query their database via API to return optimal schedules and run various “what-if” style scenarios.

Data Science

Plastics industry

15% yield increase on the mill floor due to explainable modeling techniques

Client Challenge

The client had many machine instances on the manufacturing floor. Each machine operated slightly differently through a variety of factors. Hundreds of variables were being regularly tracked in real-time. Operators wanted a visual representation of relationships between variables. Understanding causal mappings between inputs and outputs would allow them to see what actions should be taken to improve output quality.

Our Approach

Leveraging a casual inference type approach to understanding variable relationships and linear methods allowed the team to filter out uninformative features and understand directional relationships between available levers and quality. Techniques such as Granger Causality (time series), SHAP values, and PDP plots were extracted to provide both high-level and granular level insights in an easy to consume, automated report.

Outcome

Automated reports are being delivered routinely to decision makers operating on the mill floor to increase productivity. The operators now have access to what machine parameters should be adjusted in order to maximize output quality.

Data Science

CPG industry

Demand forecasting at enterprise scale for production planning shows 23% improvement

Client Challenge

Erratic demand due to Covid-19 was making business planning difficult and causing significant disruptions in the supply chain. The client needed to leverage a large amount of external information to plan production demand correctly and at a much higher granularity than previously done. Previous models only predicted production demand at the aggregate level in a rolling monthly format, but client required weekly updates at a factory level to bring adaptability in Covid struck production schedules.

Our Approach

Taking large amounts of information into a spun-up cloud VPS instance allowed parallelization across numerous vCPUs. This information was then restructured to leverage regression style approaches to time series forecasting with deep learning networks built in Tensorflow. Containerization in docker and model versioning and hosting via API allowed rapid scoring and training while minimizing compute costs.

Outcome

The comparison between the new Covid-19 robust model and the spreadsheet-based model previously used showed a 23% reduction in MAPE and levels of detail not previously possible. This allowed individual plants to generate optimal plans for production inventories.