Current
AI Engineer
Confidential · NDA
Company and project details are confidential.
Kraków, PolandSoftware developer · WaveCut
A developer by trade.
Curious about everything AI.
I’ve spent two decades building software, from backend services to products people use every day. Lately, AI has become my favorite place to explore: running models locally, training and modifying them, and bringing experiments to life in my pet projects.
Built across startups
and global products
LyftUnityRakuten Viber
The foundation
Two decades in software.
Consumer products, geospatial search, fintech and AI. New products and mature codebases, with the same responsibility for what ships.
View / print CV2025–Present
Mobiways
Founding Engineer
2023
Unity
Senior Software Engineer · B2B engagement
Go service optimization, refactoring and production reliability.
2019–2021
Lyft
Senior Software Engineer · Mapping & Search
Search services, public geospatial APIs, operational tooling and data ingestion. Contributed to a broader cost-reduction effort expected to exceed $10M.
2019
SoftSwiss
Senior Software Engineer
Real-time Go services and bare-metal Kubernetes for sports betting.
2018–2019
VIV Dev
Senior Software Engineer
Backend architecture, delivery and monitoring for a mobile health product.
2013–2018
Rakuten Viber
Senior PHP Developer
Public and private APIs behind Viber Out, Sticker Market and Viber Wallet.
2005–2013
Where it started
Web, financial and public-sector systems. From developer to senior engineer and team lead.
Selected work
Projects & experiments
Independent projects and experiments: useful software, local AI, and the occasional deep dive into how a model works.
Independent project · Since 2011
Plotva
A conversational chatbot with a life of its own.
Explore the sourceI started Plotva in 2011 as a conversational chatbot. It moved to Telegram in 2016 and grew into a community character with memory, image generation, music and its own personality.
It puts local models to work through serving, queues, retrieval and privacy filtering. The open-source runtime is now Rust, following PHP and Go generations.
Telegram conversation
Memory + contextPlotvaLocal models + tools
TextImagesMusic
Neural rendering · Apple Silicon
MLX-DLSS
I rebuilt the DLSS neural rendering and frame-generation networks for Apple Silicon with MLX and Metal, alongside a PyTorch path. The work reaches down into precision, rounding, motion and the behavior of individual kernels.
The code is open source. Models are extracted from the user’s own DLSS libraries; proprietary weights are not redistributed.
Frames + motion→
Neural rendering
MLX / MetalPyTorch
Model recovery · Kernel parity · Temporal processing
Quantization · CUDA & Metal
OrbitQuant
I built a practical implementation of the OrbitQuant research method, with Transformers and Diffusers integrations, packed inference kernels and ComfyUI nodes.
Krea 2 Turbo · Measured component storage
Text encoder + diffusion transformer. Sensitive projections remain in BF16. The model card includes paired images, latency and VRAM measurements.
Local models · A continuing experiment
How much smaller
can a model get?
I experiment with expert merging, pruning and quantization for Qwen and DeepSeek, and publish builds for local runtimes. I measure what survives the process, including capabilities that coding benchmarks can miss.
What the experiments taught meOriginal experts
Merge / prune→
Smaller model
Then re-evaluate capability.
Conceptual view · not a compression ratio
Hugging Face / WaveCut125 public model reposChecked 9 Sep 2026
Qwen · REAM-288Expert merging · BF16 source + MLX variantsDeepSeek · REAM144Pruning / merging experiments · DS4 toolchainExplore the collectionLLMs · Diffusion · Multimodal · AudioAI-assisted engineering,
with engineering discipline.
I use coding agents across research, implementation, testing and review. I define the constraints, build the feedback loops and retain responsibility for architecture and release quality.
A little beyond
models and code.
I’m also interested in audio engineering. Sound, generative media and the tools used to make them are part of the same curiosity that keeps me experimenting with AI.
More about meNotes from the work
I write down what I find.
Neural rendering
Getting DLSS running on a Mac
Rebuilding two networks, and why the rounding mattered as much as the math.
Models & experimentsHalving a Mixture-of-Experts model
Two attempts, the numbers, and why a smaller checkpoint does not mean the same model.
A project’s historyFrom a conversational chatbot to a social AI character
The long road to Plotva: communities, personality, local models and a small AI farm.
Let’s talk
Let’s build something.
A product, a technical challenge,
an AI experiment, or something we could build.