Mykola Levchuk

Über Projekt
We developed an automated B-Roll generation system that enhances static “talking head” video ads. The tool analyzes the ad script and automatically creates relevant supporting video clips featuring the product or actors interacting with it. This turns a simple one-person monologue into a dynamic and professional-looking commercial without manual video editing.

Challenge
The main challenges were technical and architectural: the low-code workflow tool couldn't handle heavy image processing or secure file management, external AI APIs were unstable and caused workflow failures, sensitive user assets required a secure handling approach, and the generated B-Roll needed to stay fully consistent with the product and the ad's narrative.

Solution
We built a hybrid architecture where n8n manages the overall workflow, while heavy AI and image-processing tasks run in custom Python services. The system includes smart retry logic to handle unstable AI APIs without breaking the flow, uses a secure AWS setup with protected media storage, and integrates advanced generative image and video models to produce realistic, context-aware product B-Rolls from text prompts.

Challenge
The client needed a system to manage and analyze household expenses across multiple properties and document types. The main challenges included extracting structured data from PDFs, images, Excel, and Word files, improving OCR accuracy for low-quality scans, and automating expense categorization using semantic search.

Solution
ProtoQ built an AI-driven document processing platform with a modular, scalable architecture.
We:
• Designed a multi-stage document pipeline (PDF → OCR → AI extraction → structured JSON).
• Integrated Google Document AI, LangChain, and OpenAI Function Calling for accurate data extraction.
• Used ChromaDB and OpenAI embeddings for intelligent expense categorization.
• Implemented an event-driven backend with Node.js, Python, and PostgreSQL for high performance.

Result:
A fast, intelligent, and scalable solution that automatically extracts, validates, and categorizes financial data from any document — turning manual accounting into a seamless, AI-powered workflow.

Über Projekt
This R&D project focused on creating high-quality digital avatars and rebuilding 3D backgrounds using video recorded from a single static camera.
The main goal was to precisely remove people or specific facial features from the video and generate a clean, stable scene using neural rendering technologies.

Challenge
The main challenge was working with video recorded from a single static camera, while most reconstruction methods require multiple viewing angles. Additionally, the video had to be processed in time-based segments, which naturally produced separate models instead of one consistent scene.
On top of that, the project required extremely precise masking - not only removing the full body from the video, but also isolating and removing specific facial features without visual artifacts.

Solution
We built a custom AI pipeline that combines advanced video matting and face parsing to accurately remove full bodies and specific facial features. After testing multiple implementations, we selected and optimized the most suitable architecture for single-camera video. We also introduced custom training logic and loss functions to better handle static backgrounds, and developed an algorithm to merge multiple video segments into one clean, consistent scene with minimal artifacts.