Context

Customer: Helthcare NGO

Industry: Medical

Problem:

An NGO, working with children, looking to optimize the way abnormal behaviours are being treated. How do we detect and capture certain gestures and provide reliable data to physicians for treatment of young patients with these behaviours?

Existing solution:

Observation by a physician – difficult to achieve full-time and to look after more than 3 patients. Able to successfully recognize 40-50% of the cases.

Our Solution

Bracelet prototype with sensors for capturing of multi-modal data, generated by movement.

Classification model for recognition, based on multi-modal fingerprints for each gesture (normal or abnormal).

Model training with data from volunteers, covering 20+ gestures.

Featured Technologies:

• Motion sensors, ToF sensors, Microphone
• Classification models

Context

Customer: Metallurgy Manufacturer

Industry: Metallurgy

Problem:

Manufacturer, operating a production line, containing moving pipes with printed codes and numbers on them, needs to keep track. How do we read and recognize these codes and numbers, rotated at a random angle on a gritty, scratched, reflecting and/or hot glowing surface?

Existing solution:

Partial solution with 3 cameras and accuracy of ~15%. No solution for hot, glowing surfaces.

Our Solution

End-to-end solution, containing sensors for image capturing, combination of AI models for processing as well as deployment of the required server infrastructure.

Integrated with customer databases and security requirements.

Featured Technologies:

• Cameras with filters and lightning
• Encoding devices
• Cascade of several Classification & Regression Models
• Infrastructure configuration

Context

Customer: Automotive Vendor

Industry: Automotive

Problem:

A business, recycling car batteries after they have been stored for long periods of time needs to ensure they are still usable. How can we measure the characteristics of such a battery without having to perform a full charge-discharge cycle?

Existing solution:

Full charge – discharge cycle, taking 12 hours. Low accuracy empyrical testing, producing high amount of false positive signals, causing reputational damage.

Our Solution

Custom AI model which classifies whether a battery meets certain criteria. Utilzing the input of a quick test, collecting 140+ battery characteristics and deploying evolution strategies and correlational analysis to reach decision.

Featured Technologies:

• Time Series Classification
• Neural Networks

Context

Customer: Manufacturer of Machines for Logistics

Industry: Logistics

Problem:

A logistics business, selling machines, needs to optimize maintenance costs for their customers. How can we evaluate warranty services prices for a period of 1-5 years to ensure they are both competitive, but still profitable?

Existing solution:

Manual work by a dedicated, specialized sales person with over 40 years of experience, holding key knowledge and difficult to replace.

Our Solution

A pricing assistant model, trained on 10 years of sales data and machine maintenance records for 40 different types of machines. Taking into consideration additinal factors, such as inflation.

Integrated with customer systems.

Featured Technologies:

• Centralized Additive Models
• Classical Machine Learning

Context

Customer: Proof of Concept

Industry: Various

Problem:

Emergency repairs of machines are 3-10 times more expensive than planned ones. How can we accurately evaluate the operational status of a machine and predict if a failure is imminent?

Existing solution:

Various solutions for different types of machines.

Our Solution

A digital twin, analysing 45 powerful statistical, dynamical and correlational features – such as electrical current, temperature, vibrations, pressure, etc. Able to predict probability chances for various types of failures.

RAG chatbot function for technical questions from staff.

Featured Technologies:

• Custom Decision Tree-based model
• Agentic LLMs

Context

Customer: Proof of Concept

Industry: Public Services

Problem:

Wildfires erupt with little warning and cause millions of damage in recent years. How can we predict 24 hours early which will allow us to save 20-40% of the wildfire area?

Existing solution:

Mathematical & Statistical algorithms

Our Solution

Using public satellite images, weather data, digital elevation model data, we developed a neural network solution and trained it.

Featured Technologies:

• Custom U-Net Convolutional Network with Attention and Diffusion heads