Over the years I’ve shipped quite a few projects – from micro-tools put together over late nights, to bigger systems that ended up in companies. Some of them stay under NDA, a few I can show off publicly.
Refio
A local AI agent for IntelliJ. Actively developed.
Refio is an IntelliJ plugin written in Kotlin, running fully offline, based on local LLM models. Three working modes (chat, plan, agent) with explicit permissions. Native function calling for 8 LLM providers (OpenAI, Anthropic, Google, Ollama, LM Studio, OpenRouter and others), with a fallback for weaker local models via XML tags. Six RAG techniques for managing big projects in limited AI memory.
I build Refio because I’m on the side of control over the process, not convenience. Cursor, Copilot, Claude Code have their own philosophy and way of doing things – Refio works according to my philosophy and my idea, full transparency for learning and figuring out how it all works under the hood. Every prompt visible, every context decision visible, every file touched – visible.
MIT license, no cloud components, JVM-auditable.
Refio comes with a local LLM benchmark at benchmark.refio.dev – a comparison of performance for agentic workflows (KV cache, token cost, latency, quality of generated code) on real consumer hardware. It’s a lab for checking which model performs how when it comes to generating code that’s supposed to work “spot-on”, and what the stability looks like across repeated runs.
Links:
- refio.dev – the project site and documentation
- benchmark.refio.dev – public benchmarks of local LLMs
- GitHub – source code
- JetBrains Marketplace – one-click install
Other projects over the years
Before that I built all sorts of things – some are alive, some are waiting “until his lordship” finds the time for them. Listed newest first.
LandingBuddy (2024-2025)

A tool for automatically generating landing pages with the help of AI. Stack: FastAPI + React + TypeScript, own infrastructure, local LLM models for generating content. Live and public: landingbuddy.com. An experimental project – a case study in the fact that distribution has to be built in parallel with the product, not after it.
ArtiMid (2024)

A local tool for generating AI graphics based on Stable Diffusion, with no queues, no logins and no external platforms. Stack: Python + TypeScript, own infrastructure. Live and public: artimid.com.
TidyPost (2024-2025)

An AI content assistant combining text generation (OpenAI, Claude) with generative graphics (SDXL) and a pipeline for publishing to social media. Stack: Python + TypeScript. Not released publicly.
PixelDream (2023-)
An internal SaaS framework – the technological foundation under LandingBuddy, ArtiMid and TidyPost. FastAPI + React 18, a system of 8 AI agents, 8 LLM providers, a model manager with imports from CivitAI and HuggingFace, an AI App Studio. Over 1190 commits since 2023. It’s a developer platform, not a product – which is why there’s no public landing page. Currently in a phase of architectural refactoring.
Stable Diffusion + prompt engineering (2023)
Experiments with generating graphics via Stable Diffusion: model comparisons, prompt engineering, my own original approach. A series of posts on the blog →
JDD 2018 talk – Introduction to DL4J
A 42-minute talk at the JDD 2018 conference (PROIDEA Events, Kraków) – neural networks from a Java developer’s perspective: from the biology of neurons, through DL4J (DeepLearning4J) as an ML library for the JVM, to my own demo app for visualizing the network training process (XOR, iris classification). The oldest documented public talk about AI in my portfolio. YouTube →
The Java 8 series (2014-2017)
A series of articles about the new features in Java 8: multithreading (4 parts), streams (java.util.stream), collections, sorting, lambdas. The most-read posts on the blog. The Java series on the blog →
“From a builder’s handbook” – design patterns (2015)
An 11-part series explaining the GoF patterns (Singleton, Observer, Mediator, Command, Composite, Decorator, Iterator, Proxy, Adapter, Facade) through construction-site analogies. The post about the Command pattern has been pulling in organic search traffic for ten years now – the blog’s bestseller. The design patterns series on the blog →
Raymarching + demoscene (2014-2015)

A raymarching shader series: distance fields, the Menger sponge, OpenGL post-processing, sky & sun, raymarching light and shadows. Plus a demo in C64 assembler at the LoadError 2015 competition. Demoscene math in its purest form. The raymarching series → · Demoscene →
S3D Rend v0.02b (2004-2005)
A 3D graphics engine written in C++ using SDL + OpenGL + OpenAL. It ran on Windows and Linux (Dev-C++, Anjuta). Archived.

s3 Shader Boy / s3 Old Media Maker / s3 Gdx Codec (2010-2014)
Three tools related to libGDX and graphics:
- s3 Shader Boy – an app for testing OpenGL 2.0 shaders on mobile devices
- s3 Old Media Maker – a multimedia editor in the style of the “RSI Demo Maker” from the Amiga 500
- s3 Gdx Codec – a Java library for saving graphics from libGDX straight to MP4 (GPU → MPEG4)
Archive.
Shad Link System (2014-2016)
An extensive PHP system for automating SEO link building: submitters, pingers, analyzers, generators, anchor management, automatic directory listing. 2-3 years of development, never released publicly – by the time I went to launch, bigger products had already taken the market.
php Painter (2007-2010)
A library in pure PHP (with GD/GD2) for image operations: thumbnails, previews, dynamic photo processing, generating CAPTCHA tokens against OCR. Archived.
Contact
Some of these projects are no longer alive, but to this day they give me experience and satisfaction. Some of the concepts come back in new projects, others stay as a good memory. If you see something interesting or want to talk over a similar project – get in touch →.