Software and hardware for radio-frequency measurements. I own the web app for controlling instruments and visualizing measurement data. Stack: TypeScript, React, TanStack, shadcn/ui, Storybook, Playwright, WebSocket, C++ and SCPI/HiSLIP.

• Designed the frontend architecture around Feature-Sliced Design, simplifying scaling and maintenance.
• Implemented client/server interaction over WebSocket and REST for exchanging data with measurement hardware.
• Integrated vector network analyzers via C++ backend logic over SCPI/HiSLIP.

A platform that streamlines collaboration between influencers and advertisers. Stack: TypeScript, React, Next.js, MongoDB, Docker, python-telegram-bot and GitLab CI.

• Led a team of 3 developers: planned tasks, ran code reviews, and owned technical delivery.
• Shipped 3 projects to production in 9 months, aligning technical decisions with business goals.
• Built a Telegram bot for influencer/advertiser interaction — 100+ ad offers processed in the first 3 months.

Analytics service for Wildberries sellers. Stack: React, TypeScript, MobX, InversifyJS, Webpack, Jest and Docker.

• Rewrote the frontend in 3 months on a new stack: React, TypeScript, MobX, and InversifyJS.
• Introduced Dependency Injection into the React app, improving testability and extensibility.
• Fixed 50+ bugs, added Jest unit tests for key modules, and mentored 2 interns.

Automation of tax and statistical reporting for companies. Stack: React, Redux, TypeScript, Next.js, MUI, Vite and Docker.

• Built a web app for electronic company statistics reporting; automation cut paperwork by 40%.
• Created an internal UI kit, speeding up typical interface development by up to 15%.
• Led a group of 3 junior developers: taught frontend fundamentals and ran code reviews.

A Telegram gateway for working with OpenClaw agents, built for a small circle of users. Stack: TypeScript, Bun, grammY, Telegram Bot API and systemd.

• On first contact, every allowed user automatically gets their own isolated agent, with its own workspace, memory, session history and reminders — no data is shared between users.
• The system identifies the type of request and routes the message to the right agent. A reminder a user creates is stored and delivered back to the same Telegram chat at the right moment.
• The architecture is split into independent components. All interaction with the OpenClaw runtime goes through a single CLI adapter, so the rest of the system does not depend on how the agents work internally.
• The core components are covered by 102 tests across 13 modules, with message routing, user-agent provisioning and the Telegram transport each tested separately.

A framework that runs ARM64 builds inside a gem5 full-system simulation on ordinary x86_64 CI agents, so cross-architecture builds need neither ARM hardware nor cloud instances. Stack: Python, gem5, Jenkins, Debian, QEMU and Bash.

• Boots a full Debian ARM64 guest under gem5 and drives it from Python: the environment is declared in YAML, an orchestrator owns the simulation lifecycle, and artifacts move over SSH.
• The guest is split across two disks — an immutable 2 GB base and a 4 GB build overlay mounted into a chroot — so toolchain updates never touch the base image and builds stay reproducible.
• Cut cycle time by 40%, from 14.6 to 8.7 minutes, by restoring gem5 checkpoints instead of cold-booting (382 s to 28 s); ccache inside the chroot took another 30–50% off incremental builds.
• Shipped as a Jenkins Shared Library, so a pipeline calls one declarative step. A month of operation logged 147 ARM64 builds at a 94.2% success rate.

An interpreter for a small programming language defined by its own context-free grammar: lexical analysis, LL(1) parsing and execution of Reverse Polish Notation. Stack: C++, Make, LL(1), RPN and BNF.

• Parses input with a table-driven LL(1) analyser that emits Reverse Polish Notation during the same pass, instead of building a tree and walking it separately.
• Executes the generated notation, covering arithmetic, conditionals, while loops and static arrays.
• The grammar is formalised up front in BNF with transition and RPN-generation tables, so the parser follows a written specification rather than ad-hoc rules.