ML / AI Engineer | Python Developer

Designed and implemented a graph-pruned hierarchical semantic index for 3D Gaussian Splatting digital twins. The index organizes scene, zone, region, object, and view evidence so queries prune irrelevant branches rather than scanning the full map.

Built a same-input graph-versus-flat evaluation testbed and an agentic retrieval pipeline using a local LLM. Automated hierarchy construction from RGB-D and CLIP, completed embedding, ranking, quality evaluation, baseline integration, and reproducible-run modules, and owned experiment configurations, traces, validators, and statistics. The work was submitted to an ECCV 2026 workshop.

Developed an autonomous real-time bot for a competitive 1v1 game and became a GAICA finalist. Implemented TCP communication with the match server, JSON message exchange, game-tick processing, and graceful degradation that returns valid commands on internal errors.

Designed tactical action planning through short game-world simulations and risk-robust worst-case evaluation. Built opponent-style recognition, combat mathematics for ballistics and shot evaluation, projectile danger maps, and an early-round weapon strategy. Trained a compact PyTorch neural policy and accelerated the hot path with a native Rust extension with a Python fallback.

Built an end-to-end semantic search and experimentation pipeline over a corpus of 500,000 documents, including data preparation, vector encoding, vector storage, approximate nearest-neighbor retrieval, and quality evaluation.

Compared lexical and neural search approaches and multiple index types, identifying encoder quality as the primary bottleneck. Selected a compressed-index and exact re-ranking trade-off that reduced memory consumption substantially with near-equivalent quality, and implemented YAML configuration, CLI workflows, caching, tests, and notebook-based demos.

Developed the backend for a production analytics dashboard for the owner of the Namaz: Learning for Beginners mobile application. Built FastAPI services with asynchronous MongoDB access, user segmentation, flexible filtering, group comparison, and data export capabilities.

Contributed to a feature that interprets plain-text data-selection filters and parses them into queries with a language model. Worked in a team environment using GitLab, merge requests, code reviews, CI/CD, architecture documentation, and release preparation for client infrastructure.

Built a service that converts free-form product descriptions into structured attributes and progressively narrows KTRU public-procurement classifier codes with explanations. Supported furniture, clothing, and tools domains.

Created an index for 3,030 KTRU catalog entries with tens of thousands of allowed values, including parsing, unit normalization, validation, and user hints. Implemented an offline explainable rule-based and fuzzy-matching pipeline with approximately 50 ms median response time, 44/45 top-1 evaluation accuracy, 136 pytest tests, FastAPI UI packaging, Docker, load measurements, and reproducible catalog builds.

Built an open-source, security-first voice assistant in which the language model recognizes intent but cannot directly control the operating system. Designed a security layer between the model and system actions.

Implemented intent parsing through rules and an LLM with strict JSON schemas, followed by default-deny policy checks, allowlists, parameter validation, decision logs, and narrowly scoped executors. Integrated OpenAI function calling, speech recognition, and speech synthesis, with mock-based unit tests isolated from the OS and external APIs.

Implemented an AI-assisted web application for task decomposition, planning, and focus management. Built the client application with React, TypeScript, and Vite, including a step-by-step goal-to-plan flow and conversational assistant.

Moved language-model requests to a server-side Node.js proxy to prevent API keys from reaching the browser. Added an offline mode based on local logic and configured cloud deployment.

Built 2D games with Pygame and desktop applications using Kivy and PyQt5. Applied object-oriented architecture, game loops, physics, and multi-screen interface design.

Also developed a deployed Telegram bot, implemented linear algebra and machine-learning algorithms from scratch, and created an evolutionary image-reconstruction algorithm using parallel computing.

Relevant coursework includes deep learning for search, machine learning, artificial intelligence, databases, distributed and network programming, computer networks, probability theory and statistics, algorithms and data structures, and operating systems.