It-Jim

The project goal was to create a mobile solution that generates accurate 2D floor plans and 3D layouts of physical spaces with real-world dimensions. By scanning rooms with a smartphone camera, the app reconstructs environments in detail, delivering both textured 3D models and precise 2D drawings.

This approach replaces time-consuming manual measurements and sketching, providing instant, professional-grade results.

Input: camera scan

Output: 2D room plan

Technologies & Skills:

• Mobile Development

• Computer Vision

• iOS Development: Swift

• ARKit

• 3D Model

• LiDAR 

• TrueDepth Sensor

Project Development Approach

We developed a computer vision pipeline that captures spatial geometry from iPhone LiDAR scans, producing both 3D models and scaled 2D floor plans. Using real-time 3D reconstruction algorithms, the system builds a detailed digital model of the scanned room.

From this model, the app automatically generates a 2D floor plan with accurate dimensions. Users can seamlessly switch between 2D and 3D views, explore the reconstructed environment, and take precise measurements.

The app workflow was designed to be simple and intuitive:

Scan → 3D Reconstruction → 2D Plan Generation → Measurement

The app functionality includes:

• Scanning the interior and result 

• 3D reconstruction model of a room with textures

• Generated 2D plan 

• Making measurements in 2D and 3D

Solution & Results

The mobile application digitizes indoor spaces in just minutes, producing ready-to-use layouts for architects, real estate professionals, and interior designers. What once took a full day of manual measuring and drawing is now reduced to a quick and automated process.

The app delivers:

• Accurate 2D floor plans with real-world dimensions

• High-quality 3D reconstructions with textures

• Instant measurement tools in both 2D and 3D views

This results in significant time savings, improved accuracy, and smoother workflows for professionals working with indoor spaces - architecture & design, real estate, facility management & construction, retail, and interior design.

The goal of the project was to evaluate a person's health vitals, such as heart rate, oxygen saturation (SpO2), and respiration rate. Mobile phone selfie camera extracted these vitals using the remote photoplethysmography (rPPG) technique.

Input: mobile camera feed

Output: heart rate, oxygen saturation, respiration rate

Technologies & Skills:

• rPPG

• Computer Vision

• Face Detection

• Signal Processing

• Programming Languages: Python, C++, Java, Swift

Project Development Approach 

To measure the heart rate, it was necessary to extract the remote photoplethysmographic (rPPG) signal from the video. This signal is an innovative combination of RGB signals (pixel values from the video frames). Pixel values, specifically color components, should be extracted from the specific regions on a face. Then, the heart rate is estimated via spectral analysis of the rPPG signal.

The solution is based on the blood volume changes in tissue due to the cardiac activity that affects the optical characteristics of reflected light. The rPPG framework included the following steps: face and ROI (region of interest) detection, facial landmarks extraction and tracking, rPPG signal extraction and processing.

Solution & Results

The implemented solutions estimate human vitals from a live camera feed.

• Custom face landmarks detector and tracker.

• Specific signal extraction procedure.

• Optimized signal processing running on the edge.

• Utilization of mobile camera feed.

• Calculation of heart rate, saturation, and respiration rate.

The project idea was to transfer selected clothing from the photos to the 3D model while preserving the necessary human pose. This solution provides added value for retailers, enabling users to shop for clothes online.

Input: 2 photos of a garment

Output: 3D model, textured mesh

Technologies & Skills: 

• Deep Learning

• Computer Vision

• Image Segmentation

• OpenCV

• PyTorch

• Open3D

• Programming language: Python

Project Development Approach

The solution is based on the pose estimation models used for landmark detection. Since clothing alters its form when taking the shape of a person’s body, for a proper augmented reality (AR) experience, a deep learning model should identify not only basic key points on the body’s joints but also the three-dimensional body shape.

The solution includes the following functionality: 

• Human clothing and body segmentation

• Dense pose estimation

• Transfer of textures from real-life photos

Solution & Results

This is an excellent solution for leveraging computer vision and AI technology in the fashion industry. The CV-enabled applications allow potential customers to preview how clothes and accessories would look on a human body, saving them time on shopping and increasing satisfaction rates. Now, retailers can engage customers with an innovative and personalized clothing experience.

The project simplifies the process of visualizing interiors, offering automatic room detection and instant style transformations from a single photo. A user can easily upload an image of a space and, with a simple text prompt, receive a high-resolution rendering of the same room redesigned in a chosen style - modern, Scandinavian, minimalist, and more - while preserving its original layout and proportions.

Input: real photo of a room (e.g., kitchen, living room, bedroom)

Output: image of the same room decorated in a particular style (Modern, Scandinavian, etc.)

Technologies & Skills:

• Computer Vision 

• Generative AI

• PyTorch

• HuggingFace Transformers

• OpenCV

• Stable Diffusion

• ControlNet

• Replicate

• Python

Project Development Approach 

In this project, we used the Stable Diffusion with Control Net technology to create an AI system for interior image manipulation. Built with Stable Diffusion and ControlNet, the system combines Generative AI and Computer Vision to ensure transformations are both creative and accurate.

It classifies room types automatically, applies one-click style changes, and even supports selective modifications to walls, furniture, or décor. Super-resolution techniques enhance image quality, while deployment on Replicate makes the solution easy to integrate through an API.

Key functionalities include:

• Changing room style in one click

• ControlNet with multiple conditional inputs

• Automatic room classification for improved output accuracy

• High-quality output with Super Resolution

Solution & Results

The system delivers photorealistic interior restyling at scale, generating new styles in seconds while preserving room geometry. With multi-style support, accuracy, and scalability, it is ready for professional and consumer applications.

Highlights: 

• Create high-quality images of any room in a new style while keeping the original layout intact

• One-click access to multiple built-in design styles

• Automatic room classification for optimized processing across different layouts

• Simple API integration and deployment on Replicate

This solution is a time and cost-saving tool for designers, real estate agents, retailers, and homeowners, enabling them to visualize spaces quickly, reduce design costs, and boost engagement.

Its applications span across interior design, real estate staging, e-commerce, renovation platforms, and AR/VR visualization, promising a more efficient and engaging process.

The goal was to design and implement a reliable autonomous driving system capable of accurately interpreting the road environment in front of a vehicle. The system analyzes road surface, lane markings, and surrounding traffic in real-time to support safer and smarter driving decisions, keeping us at the forefront of autonomous vehicle technology.

By leveraging deep learning and computer vision, it provides robust situational awareness that is essential for autonomous navigation with lane segmentation.

Input: camera feed

Output: lane masks, traffic object detection, road environment scanning+

Technologies & Skills:

• Artificial Intelligence

• Computer Vision

• Deep Learning

• Embedded AI

• Neural Network

• Programming Languages: C++, Python

Project Development Approach

We developed a custom convolutional neural network (CNN) architecture optimized for real-time inference. The network was trained to perform multiple tasks simultaneously, including detecting road surfaces, segmenting lane labels, and identifying critical traffic objects.

We trained a complex, custom neural network with multiple outputs and deployed it on the Nvidia Jetson Xavier AGX, together with code for data processing and prediction filtering in real-time. A dedicated data processing pipeline was implemented to clean incoming sensor data, while a filtering module ensured stable and reliable predictions under dynamic driving conditions.

Solution & Results

The resulting complex solution is a multi-task deep learning system that operates in real-time on embedded automotive hardware. It provides the following:

• Processes live video input and produces segmented road surfaces, lanes, and object detections simultaneously.

• Operates at real-time performance on Nvidia Jetson Xavier AGX, meeting the needs of autonomous driving applications.

• Achieves high stability through prediction filtering and data preprocessing, ensuring consistent outputs even in complex road scenarios.

This solution demonstrates how custom AI models can strike a balance between accuracy and efficiency, delivering actionable insights for safe autonomous navigation.

Value & Applications

The system provides significant value across autonomous driving and ADAS (Advanced Driver Assistance Systems) domains:

• Autonomous Vehicles: real-time road perception for navigation and safety.

• ADAS Systems: lane departure warning, adaptive cruise control, and collision avoidance.

• Smart Transportation & Mobility Platforms: enabling connected vehicles with reliable scene understanding.

• Research & Development: testing advanced multi-task AI architectures for edge deployment.

The project goal was to replace the human operator and automatically find the most interesting area on a hockey, football, or basketball pitch. The It-Jim team has built a large-scale system for real-time sports analytics.

Input: Video stream

Output: Player's positions, ROI

Technologies & Skills:

* Computer Vision

* Machine Learning

* Real-time Analytics

* ROI Control Algorithm

* Programming Language - C/C++

Project Challenges & Functionality

We used a region-of-interest (ROI) control algorithm to ensure accurate tracking of players in real time.

The system has several key features:

* Pitch mask segmentation prevents false detections of the player blobs.

* Real-time panorama stitching combines images from multiple cameras.

* Adaptive filtering and thresholding help find player candidates.

* A custom tracker accurately follows player movements.

* A unique clustering method helps with smooth and logical ROI selection and control.

Solution & Results

We have designed, developed, and deployed a large-scale system for real-time sports analytics. As a result, it was possible to replace the human operator and automatically keep track of the sports game.

The main advantage of this multi-sports tracker lies in its **high performance**. It runs at 120 FPS (frames per second) and can work in many cloud instances. This makes the solution cost-effective, flexible, and capable of real-time ROI tracking and streaming for clients worldwide.

Highlights:

* Multiple objects are tracked in real-time.

* The tracking system combines results from multiple cameras.

* The AI-based solution provides data for sports analytics.

* Automatic extraction of the region of interest for live streaming.

Crystal Clear Memories is a US-based retailer of custom 3D laser-engraved crystals. The goal was to automatically generate 3D models for volumetric laser engraving in memory crystals based on a user’s image. We developed an AI pipeline that transforms 2D images into manufacturing-ready 3D models, cutting costs, improving consistency, and accelerating delivery.

Input: photo

Output: engraved gift with 3D photo

Technologies & Skills

• 3D reconstruction

• Computer Vision

• Machine Learning

• Depth Estimation

• Image Segmentation

• 3D texturing

• Point Cloud

• Generative AI

Project Development Approach

We designed a complex AI-powered pipeline consisting of image segmentation tools, monocular depth neural networks, and computer vision algorithms.

A 'human-in-the-loop review' process was employed for handling complex cases (e.g., cat whiskers). In this process, the AI-generated 3D model is reviewed by a human operator who can make necessary adjustments to ensure the final model meets the quality standards.

AWS-based infrastructure facilitated scalability and cost efficiency. The cloud setup supported spikes in demand across multiple sales channels, significantly lowering operational costs and greatly enhancing customer satisfaction.

Solution & Results

The AI-based solution automated 85% of 3D model creation, reducing processing time from hours to minutes, with no need for 3D designer involvement. It ensured consistent, premium-quality outputs while maintaining even the most minor detail.

This automation helped to streamline the workflow and increase the number of orders while sustaining the same quality level.

Key Results:

• 85% automation of manufacturing-ready 3D model 

• Faster turnaround times for customers

• Consistent high-quality outputs with minimal human oversight

• Scalable pipeline capable of handling higher order volumes

• Lower operational costs, enabling sustainable business growth

The aim of this solution was to develop an engine that would assist with medical documentation processing, thereby reducing manual data entry time and errors. By capturing images of paper forms with a mobile device, the CV/ML engine converts them into digital documents with accurately recognized data fields (pictures, tables, text data, links, etc.).

Input: image from a mobile

Output: corrected text page

Technologies & Skills:

• Computer Vision
• Image Processing
• Automatic Document Analysis
• Automatic Document Correction
• Pattern Detection & Recognition
• Programming Languages: Python, C++

Project Development Approach

Automatic document analysis is a key part of the overall document recognition process. A typical scenario starts with taking a picture of the document with a mobile phone or tablet. Then, the service parses and recognizes content from the captured document automatically.

Since a few geometric distortions of the paper, varying illumination, and occlusions were possible, we built the page unwrapping engine, which successfully handles these issues.

The developed algorithm contains four key components: Preprocessor, Feature Extractor, Geometric Model Estimator, and Refiner.

The steps of the developed algorithm included:

• Extraction of horizontal and vertical linear spans.
• Extraction of several groups of local features.
• Document correction based on estimated parameters of the geometrical model of the page.
• Compensation of residual image distortions for both vertical and horizontal directions using a dense sampling grid.

Solution & Results

The developed Page Unwrapper solution provides functions for automated document analysis and recognition that can be implemented in a wide range of applications, including scanning, search, sorting, and information management. The created engine operates accurately and corrects distortion within a reasonable timeframe.

The document analysis and recognition solution is capable of:

• Handling illumination changes.
• Correcting nonlinear geometric distortions.
• Automatic page border detection.
• Creating a well-refined document layout.

Indoor positioning systems (IPS) are increasingly crucial for enabling real-time navigation in GPS-denied environments such as airports, hospitals, museums, shopping malls, and warehouses. In this innovative project, we aimed to design and implement an accurate and scalable indoor navigation solution based on available mobile sensors and inexpensive Bluetooth beacons.

Input: BLE beacon signal

Output: precise user location

Technologies & Skills:

• Mobile Development

• iOS Development: Swift, Objective-C

• Mobile Device

• Signal Processing

• Sensor Fusion 

• Bluetooth Beacon

• IMU (Inertial Measurement Unit)

• C/ C++

Project Development Approach 

We have developed a positioning algorithm using inexpensive Bluetooth beacons and built-in IMU sensors on a mobile device. The BLE signal processing module is based on custom algorithms providing stable positioning results. In its turn, the IMU (Inertial Measurement Unit) signal processing module provides robust step detection and device heading estimation in real-time. The positioning engine combines a BLE module and IMU-based detection, fused through a proprietary algorithm to achieve stable and precise results.

The solution was implemented in C++ for high performance and integrated on both iOS and Android platforms. To simplify deployment, we also developed a complete framework for calibration, setup, and system management.

Solution & Results 

The solution provides 1-2 meter accuracy in real time, making it suitable for a broad range of navigation and asset-tracking applications.

As a result, the system operates reliably in both wide open spaces and narrow indoor corridors, achieving real-time precision suitable for dynamic navigation. Importantly, it has already been successfully deployed in several live locations, proving its robustness and scalability. 

• Proprietary positioning engine delivering stable 1-2 m precision.

• Provides a complete framework for calibration, system setup, and management.

• Operates effectively in both open spaces and narrow corridors.

• Successfully deployed and tested across multiple locations.

This technology provides significant value in navigation, asset tracking, and customer experience enhancement, with practical applications across transportation hubs, healthcare, retail, education, and industrial settings.

The project goal was to create a system that automatically generates personalized, realistic, and stylized avatars from user-provided photos using artificial intelligence. The Avatars project enables users to transform ordinary photos into customizable AI-generated avatars, enhancing user experience and engagement across social platforms.

Input: one or several personal photos

Output: custom stylized avatar image

Technologies & Skills: 

• Artificial Intelligence

• Deep Learning

• Generative AI

• Computer Vision

• Image Processing

• Fase Swap 

Project Development Approach

The project was designed to deliver high-quality, personalized avatars at scale. Our team built a robust, automated pipeline that evolves with user needs and emerging AI capabilities.

To guarantee professional results, we implemented quality controls that dramatically reduced visual errors. We began with LoRA (Low-Rank Adaptation) training, a technique that efficiently fine-tunes large, pre-trained AI models for image generation. LoRA helped balance model size, quality, and training speed. As the project progressed, we transitioned to InstantID-powered face inpainting, enabling avatars that preserve user identity while achieving professional-grade resolution and style diversity.

Beyond static avatars, we explored style transfer and human animation, opening new opportunities for user engagement and app monetization. The result is a scalable, future-ready personalization engine that helps products stand out in the competitive digital content market.

Solution & Results

As a result, the system preserves a fully recognizable face from the input photos and offers a wide range of image styles, including both photorealistic and cartoonish options. Additionally, the solution is scalable, offering fast and accurate results, and is designed to unlock new opportunities for engagement and monetization within the application.

Key Results:

• 20% faster image generation for lower costs and smoother user experience

• 90% reduction in visual artifacts, ensuring professional-grade outputs

• Scalable cloud pipeline to handle growing user demand

• Future-ready features (style transfer & human animation) for new revenue streams

Pattern recognition is an extensive field of research since there are a number of things to detect and recognize from the real world. In this project, our goal was to develop an engine for the automatic counting of the various symbols on scanned building plans. Our custom algorithm was able to handle hundreds of different symbols, including complex ones consisting of multiple small sub-patterns and text. It is robust to rotation, scaling and partial symbol occlusions. The developed engine is based on a purely classic computer vision algorithm, works very fast and does not require GPU, which makes it very cheap to use in the cloud. In addition, a comparative test with market competitors indicates an advantage from both accuracy and performance points of view.

KEY POINTS:

1. Automatic recognition of symbols with different complexity
2. Optimized search using multi-threading
3. Custom decomposition algorithm
4. Rotation and scale-invariant recognition engine
5. Successfully integrated into an engineering platform

Case study

The goal of this project was to develop and optimize the image detection and tracking algorithm for AR applications. The main challenge was to make it work directly in the mobile web front-end with all computations done on the edge.
To avoid dependence on internet connection and potential lags, we have built an efficient data processing pipeline for web front-end deployment. Firstly, a custom C++ code was developed for content-based image retrieval and tracking. Secondly, we modified and recompiled the C++ code to the WebAssembly binary code using Emscripten SDK to run it directly in the browser. Finally, the algorithm was fine-tuned and optimized within the Emscripten to ensure proper performance on the web.
Case study