I build thoughtful, reliable software for people and businesses.

The challenge

The installation needed four AI speakers to feel visually and acoustically synchronized in real time. Managing voice input, keyboard controls, API communication, and per-persona state — all within a fixed thesis timeline — required careful architecture from the start.

What I did

I built the React frontend using Vite and TypeScript, designing isolated components for each philosopher while letting a shared controller coordinate the debate. Voice input ran through the Web Speech API, an operator panel gave supervisors control over the installation, and the core interaction logic was covered with Vitest.

The outcome

The installation ran live at the thesis presentation without issues. The clean frontend/backend boundary meant the AI team could swap models without touching my code — my clearest example of building a complex real-time interface under a fixed deadline.

The challenge

Sweden's congestion tax rules involve several interacting constraints: vehicle exemptions, toll-free dates, time-of-day pricing bands, and a single-charge rule that caps fees within a 60-minute window. Encoding all of this correctly — and proving it with automated tests — was the core challenge.

What I did

I modeled the domain in a framework-free C# layer so the rules could be verified in isolation with xUnit. An ASP.NET Core API sits on top, and a Next.js frontend lets users select a vehicle type and date range to see the calculated fee immediately.

The outcome

Every business rule is tested at the unit level, keeping the API and UI thin and easy to change. Adding a new vehicle type or adjusting a fee requires touching one layer only.

The challenge

The goal was a working CLI tool that demonstrated layered architecture and dependency inversion without becoming over-engineered. It had to be genuinely installable and maintainable, not just a diagram.

What I did

I separated the project into three layers — CLI, service, and repository — connected by interfaces so each can be swapped independently. Tasks use UUID identity, persistence uses JSON with backward-compatible loading, and a GitHub Actions pipeline runs tests and publishes to PyPI on each release.

The outcome

Taskforge is live on PyPI and installable with pip. The CI pipeline runs automatically on every push, and the architecture serves as a concrete reference for clean separation of concerns in Python without unnecessary complexity.