Discover our work
Software solutions we have delivered. We work hard and we are dedicated to producing the best possible outcome for our clients, as you can see from our case studies.
Contact us for the same client experience
Sales and upsell prediction modeling
In this project we analyzed customer database, purchase history, renewal history, and CLSV to find the customers that are the most likely to purchase or upgrade. This reduced churn and shortened sales funnel and time to upgrade. Also, we looked at what made customers buy in the past to repeat the same process with new customers and make existing customers buy again. Through analyzing available data, we built predictive models that help shorten the time to convert and increase conversion rates.
Generating 3D Animation for Games, AR & VR Solutions
Solution for automatic generation of 3D objects animation for various purposes including AR & VR applications and 3D games. AI-driven technology allows to reproduce a wide variety of natural-looking virtual objects like cars, characters, etc, including interactions between a characters and an objects. The main benefit of this solution is that through automation we drastically simplify and lower cost of 3D objects and character animation while supporting wide range of object modification.
Product recognition for vending machines/smart fridges
We can build a computer vision based app for real-time detection and recognition of products from video. The solution can be used in stores to transform small self-serving shops into fully automated Amazon Go-like cashierless stores. It can help food delivery services track dispatched orders. For brick and mortar stores it can help streamline the purchasing process, reduces shoplifting, and cut costs on RFID tags.
Ecommerce recommendation engine
We’ve built a custom product recommendation system for one of the largest social platform for shopping. The system allows our client to recommend the right products and increases upsell and cross-sell opportunities, without relying on generic external systems. A complex ensemble of trained ML models include collaborative filtering models, natural language processing using deep recurrent neural networks and image processing using deep convolutional neural networks.
Retail Traffic Analysis (Visualytics Store)
We can deliver visitor analytics solutions that allow physical stores to collect and analyze in-store footfall data. We employ Computer Vision and AI to help brick-and-mortar retailers effectively manage store space and capitalize on customer insights gathered by camera-based people counting. Main features: people counting and visitor pathmaps building, visitor heatmap generation: crowd count heatmaps, dwell time heatmaps, overall traffic heatmaps, Indoor wayfinding (draw rate) assessment, queue monitoring and management, MAG (mood/age/gender) visitor profiling.
ATM fraud detection (Cook Security Group)
The engine was written in Python and C++.
If someone stays long in front of the ATM, the camera detects this event and creates a bookmark including timestamp, duration, etc.Then this bookmark is used to analyze the transaction. If there is no transaction, this means that he/she who stands in front of the ATM is not performing transactions. The YOLOv3 model is used to detect humans. For the camera video stream, OpenCV FFMPEG was used.
AI-powered attention heatmap
Perceptbox is a visual content analytics platform that enables to effectively foresee human's attention and attraction. It simulates human vision during the first few seconds of exposure to visuals, and creates an attention heatmap based on an AI algorithms that predicts what a real human would be most likely to look at. Perceptbox can identify which elements of visual content are being looked at and which are being ignored. The app is written in C++ and we've used an OpenCV and ANN.
Forex prediction system
Since the price dataset is chaotic time-series data, we,ve decided to use a machine learning model to predict the next possible EUR/USD rate. As you know, the EUR/USD market occupies over 95% of the transactions in the Forex market. We've tried to use the LSTM-CNN model at first but finally we've used ESN (Echo State Network). For the backtest, the historical data from Truefx.com was used. Truefx.com supports a very good Python API so that we can get both historical data and live feed according to our demands.
Social distance measurement
We've built models which can measure the distance between pedestrians in public places. Using a traffic light indicator system, the algorithm is able to anonymously identify and label people who maintain safe distances, while flagging certain instances in red where social distancing measures are violated. Using this information, it is possible to identify bottlenecks where social distancing cannot be maintained, and how citizens adapt as restrictions are imposed or lifted.
Gender, age, emotion detection
We've used Tensorflow and OpenCV to built this application. IMDB-2013 dataset and UTKFace dataset are used to train the custom model.This can be also used for eye color detection projects since it extracts eyes from a given image.
Intelligent lead scoring
In this project we’ve automated lead scoring and enriched historical data. The trained Machine Learning model automatically qualifies and scores all existing and new coming leads based on client’s preferred criteria like lead source, location, level of interaction or interests.
Data extraction from invoices, forms & other unstructured documents
We've have built data extraction tool that retrieves information in the key-value format and transforms documents into business-ready data better prepared for processing, analysis, and storage.
Weapon detection (AiPOD 2)
In this project, the machine learning model was built to detect various weapons including AK47, AR15, M4, Glock900, and Sigsauer. To train the custom model, a dataset which has 10,000 images with annotation information is used. The solution is based on the YOLOv3 model. (YOLO - You Only Look Once).
In this project, a PyTorch model based on Unet with MobileNet v2 backend is used for hair segmentation.
Figaro1k dataset is used to train the model.