Hi, I'm Stefan

A versatile software developer who architects performant, end-to-end systems from zero to one.

Rust  C  Python  JS/TS  Swift  Go  SQL  HTML  CSS

System Design

Engineering elegant and performant systems by resolving core architectural trade-offs.

Full-Stack

Developing across the stack, from custom frontends to high-performance backend APIs.

Data Visualization

Transforming complex datasets into intuitive and interactive 3D visualizations.

AI Automation

Automating key processes and workflows using cutting-edge AI models.

Data Pipelines

Building efficient and reliable ETL/ELT pipelines for large-scale data processing.

DevOps

Automating deployment, scaling, and operations with CI/CD and containerization.

About

I am a systems thinker who ships products to real users, driving engagement and solving complex problems with efficient code. For the past four years, I have been creating and deploying production-grade software in AI, data engineering, and high-performance applications.

At The Idea Compiler, my first full-time role, I became lead developer. While running a small team, I designed and built GyftIQ, from the React frontend to the AWS infrastructure to data analytics. Later I created Maska so agents could test it the way real people would.

I have been programming since 2012, so my computer science foundation was set before the AI era, and I have been experimenting with LLMs since late 2022. My projects in voice (Otis), shell (Giverny), artifacts (Histos), search (GyftIQ), and swarms (Maska) put me at the forefront of how humans, AI, and computers interface with each other. I work every layer, from bit shifts to devops to product management, and I know how to choose the right tool for the job, from low-level C and Rust for data-intensive pipelines to high-level Python and JavaScript for AI and web services.

While building these systems, I drew on five years of esports coaching experience where I learned both team dynamics and strategic system design. Crafting competitive tactics taught me to analyze complex systems, identify critical paths, and optimize under constraints. I apply the same thinking to technical architecture and product design. Working with teams taught me how different personalities mesh, how to communicate across skill gaps, and how to be collaborative under pressure.

The projects showcased here are examples of this approach. My process begins by identifying a challenge, analyzing its core tensions, and architecting a complete, optimized system from first principles to resolve them. This philosophy is what drives me: end-to-end system ownership, from its initial conceptual design to its deployment and continuous iteration.

Works

gyftiq

GyftIQ

An AI powered gift finder algorithm.

An AI powered gift finder that works like a collaborator, not a chatbot. You describe the person, it suggests real products, and every like or change of direction shapes the next recommendation.

React Native Expo Python AWS Lambda PostgreSQL Terraform OpenAI API
maska

Maska

A swarm of simulated users for any web app.

Turns social profiles into a swarm of browser agents that use a target webapp the way those people would. Teams get the analytics of a live launch for a fraction of the cost spent on generating real traffic.

TypeScript Playwright LLM Agents oh-my-pi
otis

Otis

Speak commands straight into your terminal.

Otis is a dictation daemon built for macOS. The defining feature that sets it apart from other ASR apps is the command mode, which lets you speak any programming or shell native language directly into your terminal. Say "pipe" and you get |.

Swift
histos

Histos

Weave an entire web app into one HTML file.

A tool for agents and engineers to weave webapp assets into a self-contained html file. Specifically designed with WASM in mind, binaries get brotli-compressed, base64-encoded and stored in pre tags. Modern AI agents have a difficult time editing massive single files, a modularized approach is superior. Histos unlocks the best of both worlds. This tool is used to build my Mecha game, Paleomap3D and the hockey visualization.

Rust Javascript
giverny

Giverny

The agent harness melts into the shell.

A composable agentic shell that dissolves an agent harness into your terminal. Invoke an LLM as a first-class shell program via a single-character alias (? do thing) that pipes, chains, and redirects like any Unix tool. Giverny was built around the core insight that the shell is the agent harness, so it skips bespoke JSON tool schemas in favor of the native commands a model is pre-trained on.

Typescript Bun OpenAI API CLI/TUI
DOOM.html

DOOM HTML

DOOM, as one utf-8 file.

For a while now you would hear that someone got DOOM running on things that should not run DOOM. This version is a playable UTF-8 string: Chocolate Doom compiled to WebAssembly, WADs and all, packed by Histos into one HTML file. Email a friend, no server, no install, no network.

C WASM Emscripten JavaScript
Mecha

Early Stage Mecha Game

An arena for testing game mechanics.

A work in progress, meant to be a endless arena mode to test and iterate upon game mechanics before applying assets such as game art and sounds. Packs down to a single 300kb file using Histos. Built using C and Raylib using a "big struct" style of code.

C Raylib WASM Emscripten
paleomap3D

PaleoMap3D

500 million years of continental drift in 3D.

An interactive 3D globe visualizing continental drift and paleogeographic data across 500+ million years. Built with Three.js on the client-side and a custom data pipeline written in C, Rust, and Python that processes large NetCDF datasets into a compressed binary format or texture for efficient delivery.

JS TS Three.js HTML/CSS Bun C Rust Python Docker
pokepastefix

Pokepastefix & HQmons

Fixes a community site for 1500+ daily users.

A client-server solution to fix and enhance missing functionality of the pokepast.es website. The browser extension replaces missing images using dynamic inline resource insertion, while the art is served by a custom Bun and Elysia.js backend. The extension serves over 1500 daily active users, with the API handling 80k+ monthly requests.

JavaScript TypeScript Bun Docker Browser Extension HTML/CSS
ow2-replaycode-ocr

OW2 Replaycode OCR

OCR on replay codes from screenshots.

A high-accuracy Discord bot for extracting Overwatch 2 replay codes from screenshots. It uses OpenCV for template matching and a dual-pipeline approach with Tesseract-OCR and the Granite Vision VLM for character recognition. Deployed on over 150 servers and used by professional teams like GenG.

Python AI/ML OpenCV PyTorch PostgreSQL Docker Tesseract-OCR
deadlocker.tv

Deadlocker.tv

A tactical map and toolkit for Deadlock.

A community hub for the game Deadlock, featuring a suite of custom-built tools, including an interactive tactical map for strategy planning. Built with a high-performance Go backend using templ for server-side rendering and a vanilla JavaScript frontend leveraging the Browser Canvas API for real-time drawing and interaction. The entire application is containerized with Docker for seamless deployment.

JavaScript HTML/CSS Browser Canvas Go templ Docker
hockey3d

NHL Play-by-play 3D

Replays NHL games in a 3D rink.

A 3D visualization tool that reconstructs NHL games by plotting play-by-play event data in a virtual rink. The project features a Rust-based parser that fetches and processes official NHL API data into a structured format. This data is then rendered on the client-side using JavaScript and Three.js, creating an interactive 3D representation of the game flow directly in the browser with HTML/CSS.

JavaScript HTML/CSS Three.js Rust

Contact

Have a question or want to work together? Feel free to reach out.