We build AI systems that
solve real problems.

HubLab's flagship product. Users describe a Telegram bot in plain language — the platform generates production-ready code with multi-step dialogues, CRM integrations, and payment modules. Fully automated CI/CD pipeline deploys and manages bots in real-time. Built from scratch with a proprietary code generation engine.

Long-term engagement with a European travel marketplace specializing in cruise and ferry bookings. Built an LLM Router that classifies inbound customer messages and routes them with CRM memory for context-aware responses. Full cycle: from research model on Jupyter/CLI to Dockerized backend with Celery workers and REST API. Event-driven pipeline processes thousands of messages daily.

Fractional CTO role at a legal AI platform. Led a 9-person engineering team through product-market fit, built the analytics pipeline tracking user behavior across 48,000+ accounts, and developed an AI-based user classification system that segments users by engagement patterns. Delivered architectural decisions on scalability and data strategy. Client under NDA.

Deep Research Agentic pipeline for a US LegalTech startup that raised a $1M seed round. The system automates patent database analysis using a multi-agent architecture: one agent parses legal constraints, another performs semantic search across patent databases, and a coordinator synthesizes results into structured reports. Designed for patent attorneys who need comprehensive prior art research.

One of the earliest LLM-powered tools built on ChatGPT API in 2022. Users upload NDAs and contracts — the system highlights risk clauses, identifies unfair terms, flags jurisdiction mismatches, and provides a risk score. Designed for startups and freelancers who need quick legal review without hiring a lawyer. Later expanded to support multi-language contracts.

Real-time computer vision system for detecting exam violations from overhead cameras. Trained two model families (YOLO and DETR) on custom-annotated datasets of prohibited objects and suspicious behavior. Includes camera distortion calibration, multi-object tracking, and a dashboard for invigilators with instant alerts. Deployed on-premise in exam centers.

Mobile application for early screening of age-related cognitive conditions through hand gesture tracking. Integrated Google MediaPipe for real-time hand landmark detection with a custom ML classifier trained on 7,000+ annotated gesture images. Achieved +11% accuracy improvement over baseline. Won «Best Social Impact» at AstanaHub Battle 2023 ($25K award from Tim Draper). Accepted into Draper University accelerator (USA).

Computer vision defect segmentation model for quality control in 3D printing of spacecraft structural components. Led a team of 12+ engineers. The startup was accepted into the Creative Destruction Lab (CDL) Space Stream at the University of Toronto — one of the top space-tech accelerators globally. Model deployed on-premise directly on the manufacturing device.

Enterprise chatbot built on Retrieval-Augmented Generation for a freight transportation company. Documents are chunked, embedded into vectors, and stored in a vector database. Employees type natural language questions and receive precise answers with source references. Reduced average document lookup time and improved onboarding speed for new staff.

Telegram-based AI sales assistant for a hotel chain in Thailand. Built custom parsers for three major hotel aggregator platforms (Booking, Agoda, Hotels.com) to provide real-time price comparisons. The bot handles inquiries 24/7, suggesting rooms, answering FAQ, and routing qualified leads to human managers. International remote team, solo backend development.

Proof of Concept analytics platform for a UK school network. Combines RAG for qualitative queries with Text-to-SQL for quantitative questions. Teachers and administrators type questions like «average attendance last month» and the system automatically generates SQL queries, retrieves data, and presents visualized answers. Hybrid AI approach handling both structured and unstructured school data.

Published research on a novel data augmentation method for thermal imaging. Developed a technique that mimics visible-spectrum characteristics in infrared images to improve object detection for autonomous driving. Achieved ~13% improvement in prediction accuracy over baseline models. Research conducted and published at ITMO University, trained on PyTorch with custom augmentation pipelines.