Wiktor Zoga

I build software and explore AI systems, machine learning, and performance.

I learn by building: turning ideas into small working systems, testing them, and understanding the trade-offs behind them.

More about me

Projects

View projects

In progress

A GPT built from scratch in PyTorch, with experiments on how model size and attention affect runtime and memory.

A learning project, still in progress.

GitHub

In progress · PyTorch

LLM Systems Lab

My learning project: building a GPT and measuring how its implementation affects time and memory.

What’s there

A GPT implementation, training and text generation, GPT-2 weight loading, and synthetic benchmarks. The experiments compare model depth, memory use, and attention implementations.

What’s next

KV caching and inference benchmarks. The work is unfinished, and I’m still exploring the design and measurements.

View on GitHub

What I'm exploring

AI systems

Agents, tools, and the architecture behind useful AI products.

Machine learning

Building models and testing their behaviour, starting with GPT in PyTorch.

Performance

Understanding runtime, memory, and the systems beneath the code.

Writing

Notes from building and learning. No posts yet.

Blog