MC.

Software developer, based in Graz

01 / Projects

Projects

THINK AHEAD!

A website for a nonprofit educational board game from Graz that helps young people prepare for life after school, covering education, finances, health, and housing.

  • Nuxt
  • Vue
  • TypeScript
  • UnoCSS
  • Nitro
  • Umami
View project details
Dark retro preview of THINK AHEAD! with a charcoal window, a simple game board, and question cards.Cream and burnt-orange illustration of THINK AHEAD!.

My contribution

Built the responsive Nuxt/Vue website with TypeScript, reusable UI components, typed content models, and UnoCSS design tokens. Implemented a data-driven quiz with difficulty levels, randomized questions, scoring, and keyboard focus management. Added consent-aware Umami event tracking through a server-side Nitro endpoint with payload validation and privacy-signal checks. Configured route-specific SEO metadata, canonical URLs, structured data, and XML sitemap generation.

Technical outcome

A component-based website that brings the game, quiz, and downloadable materials together. Optimized web performance with compressed WebP assets and self-hosted variable fonts; improved usability and accessibility with responsive layouts, semantic markup, keyboard navigation, and reduced-motion support.

Open project

ALiVE — Dialogues Across Time

A locally hosted research prototype for immersive history education, developed for my master’s thesis at TU Graz. ALiVE lets users create historical figures from uploaded documents and converse with them through source-referenced answers, synthesized speech, and animated 3D avatars.

  • Python
  • Flask
  • LangChain
  • FAISS
  • llama.cpp
  • Next.js
  • React
  • TypeScript
  • drei
  • Coqui TTS
  • Docker
  • RAGAS
View project details
Dark retro illustration of ALiVE with a charcoal window, a pixel-style Einstein avatar, a conversation, and source documents.Cream and burnt-orange illustration of ALiVE.

My contribution

Built a Python/Flask REST API with 17 endpoints and a TypeScript frontend using Next.js and React. Implemented PDF/text ingestion, configurable chunking and embeddings, and persisted FAISS indexes per character and retrieval configuration. Integrated local Mistral 7B inference through llama.cpp alongside an optional OpenAI backend, with conversational retrieval and persona prompts in LangChain. Connected Coqui TTS audio playback to idle/talking animations for glTF avatars using drei.

Technical outcome

Delivered a Dockerized application connecting document ingestion, LLM inference, speech synthesis, and 3D rendering. Reusable vector indexes avoid repeated document preprocessing, while adjustable retrieval settings and visible source passages make the RAG workflow inspectable. Evaluated retrieval and answer quality through an automated RAGAS benchmark.

Read thesis

Molecular Structure Classification

A three-person university project exploring machine learning for molecular structure classification. We trained models on different representations of molecules: rendered structure images, Morgan fingerprints, and numerical 2D/3D descriptors derived from SMILES strings. The project compares image-based and feature-based approaches, including dimensionality reduction with PCA.

  • Python
  • RDKit
  • Mordred
  • TensorFlow
  • Keras
  • scikit-learn
  • pandas
  • NumPy
View project details
Dark retro illustration of molecular structure classification with a molecule, a neural network, and class prediction bars in a charcoal window.Cream and burnt-orange illustration of molecular structure classification.

My contribution

Implemented SMILES preprocessing, molecular image generation, and 2D/3D descriptor extraction with RDKit and Mordred, including filtering invalid structures and descriptor errors. Developed and trained TensorFlow/Keras classifiers: an image CNN and dense networks using 512-bit Morgan fingerprints, optional molecular descriptors, and a 64-component PCA variant. Added dropout and L2 regularization, model persistence, training/prediction commands, batch prediction exports, and shell scripts to automate preprocessing and model training.

Technical outcome

Seven training configurations cover an image CNN and fingerprint networks with optional 2D/3D descriptors and PCA. Saved models, training curves, per-run metrics, and CSV prediction exports make the experiments inspectable and reusable. The workflow connects raw molecular data to feature generation, training, and inference on new structures.

FindMyGenre

A machine-learning music-discovery app developed by our three-person university team. We trained and compared three neural-network approaches on Free Music Archive audio features to classify recordings into 16 genres. A React Native app lets users record music, compare the models’ predictions, and discover similar tracks.

  • React Native
  • Expo
  • JavaScript
  • Python
  • Flask
  • pandas
  • PyTorch
  • TensorFlow
  • librosa
View project details
Dark retro illustration of FindMyGenre with a phone recording audio, a jazz genre prediction, and suggested tracks.Cream and burnt-orange illustration of FindMyGenre.

My contribution

Built the React Native/Expo frontend for the team’s ML pipeline, with reusable screens and components, navigation, microphone permissions, audio recording and playback, and recording-state management. Connected recorded audio to the Flask inference API and implemented the interface for comparing predictions from three classifiers. Built the genre lookup endpoint over FMA metadata and integrated similar-track results, loading states, and external song links.

Technical outcome

Trained a fully connected network and a 2D CNN from scratch in PyTorch, and fine-tuned an ImageNet-pretrained DenseNet201 in TensorFlow/Keras on MFCC audio features. The training pipeline uses stratified train/test splits and class-balanced resampling, with accuracy/loss curves and confusion matrices for evaluation. Integrated these team-trained models into the recording-to-prediction workflow, alongside similar-track retrieval using cosine similarity.

Beyond Catastrophe

A top-down pixel-art survival game developed by our five-person university team in Unity. Set in a world shaped by climate catastrophe and resource conflict, the prototype combines exploration, item management, and survival mechanics influenced by day/night cycles, seasons, and temperature.

  • Unity
  • C#
  • 2D Physics
  • Tilemaps
  • Unity UI
  • WebGL
View project details
Dark retro illustration of Beyond Catastrophe with a top-down pixel-art survivor, a shelter, winter landscape, and survival status bars.Cream and burnt-orange illustration of Beyond Catastrophe.

My contribution

Co-managed the project and co-developed its C#/Unity gameplay systems. Worked across player movement and directional sprites, camera follow, item and inventory interactions, and player-status mechanics. Helped connect scene transitions, persistent scene state, and environmental systems into the gameplay loop, while coordinating tasks and development milestones with the team.

Technical outcome

A Unity prototype connecting exploration with inventory, consumables, equipment, sleeping, and temperature-dependent survival. Additive scene loading supports movement between areas, while day/night and seasonal systems influence lighting and environmental temperature. The team also produced a game design document covering mechanics, narrative, accessibility goals, and development planning.

CHC Verification Tool

A command-line tool for formal program verification, developed for my bachelor’s thesis at TU Graz. It translates programs written in a Boogie-like language into Constrained Horn Clauses and uses Z3 to check whether their assertions hold under the given assumptions.

  • Scala
  • Parser combinators
  • Python
  • Z3
  • CHCs
View project details
Dark retro illustration of the CHC Verification Tool with a code-to-clauses-to-Z3 pipeline and a terminal displaying a satisfiable result.Cream and burnt-orange illustration of CHC Verification Tool.

My contribution

Implemented the Scala parser with parser combinators and case classes representing the abstract syntax tree. Built the recursive weakest liberal precondition transformation for assignments, branches, loops, and procedure calls. Implemented clause simplification and scope-aware variable renaming to normalize the generated formulas. Generated and executed Python code using the Z3 API, including declarations for predicates and loop invariants, with support for integer and configurable bit-vector arithmetic.

Technical outcome

An automated pipeline from source text to solver results, with intermediate transformations available for inspection. The tool reports sat, unsat, or unknown and prints models and inferred loop invariants when available. Checked its behavior with valid and invalid example programs and investigated solver runtime limitations.

Charcoal portfolio illustration with a desktop window, portrait, sidebar, career panels, and command terminal.Cream and burnt-orange portfolio illustration with a desktop window, portrait, sidebar, career panels, and command terminal.

Portfolio — Retro Edition

A bilingual portfolio with a minimal landing page and a retro interface inspired by classic desktop applications. Includes animated SVG previews, a command terminal, and CRT transitions, with automated browser and integration tests.

  • Nuxt
  • Vue
  • TypeScript
  • UnoCSS
  • i18n
  • SVG
  • View Transitions
  • Playwright
  • node:test
View project details

My contribution

Built both interfaces with Nuxt, Vue, TypeScript, and UnoCSS, using shared content models, English/German translations, and reusable preview components. Created animated SVG illustrations, screenshot card flips, the design inspector, and View Transition effects. Added Playwright tests for Chromium and WebKit, plus Node.js unit and HTTP integration tests covering routing, SEO, preferences, and clipboard behavior.

Technical outcome

Content is maintained once and rendered in two responsive interfaces. Automated checks cover navigation, dialogs, keyboard focus, touch interactions, saved theme and language preferences, reduced-motion behavior, and preview playback. Shared components keep interactions consistent across both layouts, while server-rendered metadata and localized routes make each page directly accessible.

02 / Publications

Publications

Immersive learning in history education: Exploring the capabilities of virtual avatars and large language models

Steinmaurer, A., Dengel, A., Comanici, M., Buchner, J., Memminger, J., & Gütl, C.

Winner of Best Short Paper Award

View publication details
Dark retro pixel art of a research paper, neural network, and gold trophy with the text: Winner of Best Short Paper Award.Cream and burnt-orange research illustration with a paper, neural network, and Best Short Paper Award trophy.

Steinmaurer, A., Dengel, A., Comanici, M., Buchner, J., Memminger, J., & Gütl, C. (2025). Immersive learning in history education: Exploring the capabilities of virtual avatars and large language models. In J. M. Krüger, D. Pedrosa, D. Beck, M.-L. Bourguet, A. Dengel, R. Ghannam, A. Miller, A. Peña-Rios, & J. Richter (Eds.), Immersive Learning Research Network – 10th International Conference on Immersive Learning, iLRN 2024, Revised Selected Papers (pp. 363–374). Communications in Computer and Information Science (Vol. 2271). Springer Science and Business Media Deutschland GmbH.

Open DOI

03 / Contact

Get in touch.

Available for projects Download CV