I build tools to study how models learn, and software that turns ideas into working systems. These showcases introduce the projects and my contributions, with diagrams and links to explore further.

Explore the projects

LLM pre-training · Research tooling

LucidGPT

GPT-style pre-training with simple launch commands, multiple optimizer modes, built-in benchmark evaluation, and optimizer/activation monitors for training stability.

Explore LucidGPT → LucidGPT source ↗

Software engineering · Communication

What-s-up

A chat software project. My contributions covered backend communication, functional testing, and debugging: making the communication behavior work and checking it carefully.

Explore What-s-up → What-s-up source ↗

AI4Game · Open notebook

AI4Game

A place to document research questions, experiment design, implementation decisions, and lessons. The current entry is an outline; technical details and results are still to come.

Open the project notebook → All AI4Game notes →

The LucidGPT and What-s-up summaries and repository links come from my October 2026 CV. Read more about my research and background on the About page.

Inside a pre-training workflow

Launch, optimize, monitor, evaluate: a conceptual view of the capabilities described for LucidGPT. Use the steps to explore each part.

LucidGPT · capability map
LaunchSimple commandsOptimizeOptimizer modesMonitorOptimizer + activationsEvaluateBuilt-in benchmarks

Swipe the diagram sideways to see all of it →

1/0 Launch a GPT-style pre-training run with a simple command.

Conceptual capability map based on the October 2026 CV. This is an explanatory diagram, not an execution trace or measured result.

The LucidGPT showcase also includes an interactive loss-curve example. Its values are synthetic; this diagram and the curve illustrate how to explain a training run.

A workflow worth documenting

Research journey · workflow example
QuestionExperimentObservationKeep the failure. Refine the next question.

Swipe the diagram sideways to see all of it →

1/0 Start with a question and make the assumption explicit.

Illustrative workflow. A failed attempt is evidence to carry into the next iteration.

This is an illustrative research workflow, not a completed AI4Game milestone.

Tools for explaining the work

The visualization examples include interactive charts, comparisons, sortable tables, timelines, and diagrams for future research posts. Open the data view, inspect a checkpoint, or step through a process to see how a figure can make the reasoning easier to follow.