We help our partners create and implement embedded software for non-standard object detection, classification, tracking, and behavioral analysis on edge devices – such as static or vehicle-mounted cameras. For more demanding analyses, we use additional embedded computers to offload part of the tasks. A few examples of objects whose behavior we helped to analyze are airplanes at small airports, trash cans in cities, airport trolleys, cyclists, livestock…

An automotive industry client wanted to automate the process of detecting even the smallest defects in the paint coating on car parts. We developed a unique combination of custom made illumination and camera modules with in-house trained convolution neural networks. The whole solution is prepared for on-premises and cloud deployment with the industry standard in mind. We detect and classify several types of surface defects less than 0.1 mm in size with more than 95% accuracy.

A furniture aggregator platform was looking to extend their search services with visual search, which allows customers to find similar products just by uploading a photo they have taken or downloaded. We prepared the whole service as a separate microservice for the customer’s Kubernetes infrastructure. The team consisted of two ML experts and two Senior Developers, who completely designed and developed the whole service including the communication model with other already existing services. We worked with the latest architecture designs such as detector, feature vector extractor, approximate nearest neighbor and other services.

Our in-house Natural Language Processing (NLP) solution can analyse thousands of documents a day. It determines the type of content in the documents (while respecting their privacy) and can store them in a database according to predefined content types. ​For example, if you had 10,000 unlabelled documents in one folder, you could use the PDF classifier to sort them into four different folders automatically, without having to open a single PDF.

A client specialising in Facial Recognition approached us with a project on optimization of Deep Neural Network (DNN) models for TensorRT to enable high performance inference on NVIDIA GPUs. One of our tasks was finding a balance between computation load and precision for DNN quantization processes. We prepared AI video analytics pipelines with features such as object detection, classification and recognition. These pipelines can be extended by additional DNN approaches (e.g., tracking) and are based on DeepStream SDK with custom GStreamer plugins built by our company.

We created an ML engine that analyzes unusual sounds emanating from cars and provides a predictive diagnosis of the issue. It is based on unique classification technology
utilizing a custom Densely Connected Convolutional Network, delivering 98% accuracy after training on extremely small data sets of only a few sound samples for each class. The entire solution operates and employs the latest ML services from the Microsoft Azure cloud.

Our client, the owner of a private dental clinic, approached us with a request to automate patient tracking based on a series of photographs of the patient’s face and projections of their jaws taken with a smartphone. The algorithms we developed allowed us to identify some of the most important key elements used in dental diagnostics. The knowledge obtained from each photograph is combined into a single data point, and based on this our system will make a comprehensive diagnosis and offer a treatment plan. The patient can now perform these examinations remotely by himself/herself and send the results to his or her doctor.

A customer with a long term experience in AI/ML was looking for highly qualified Machine Learning developers experienced in real-time object detection and tracking, as well as behavioral analysis of objects in video feeds from a retail environment. We created an expert team comprising an ML specialist with experience of real-time object detection pipeline, two senior developers with experience in using Nvidia DeepStream SDK principles, and one annotator with experience of preparing ground truth data for ML models. Our team has substantially helped improve current customer solution in terms of performance, reliability and speed of processing.