I study Artificial Intelligence at Johannes Kepler University Linz and build AI products end to end: the model, the guardrails around it, and the app people actually touch. Most of what's below started as a 24-hour hackathon and kept going after the deadline.
What I care about in an AI system: it shows its evidence, it says "not sure" instead of guessing, and a human makes the final call.
Upload an E. coli genome, get a calibrated call per antibiotic, backed by the resistance genes found, and an honest no-call when the evidence is thin. Trained and tested on 2,127 genomes, held out by lineage.🏆 Winning project, Hack-Nation 6 (Vienna hub) · Python · scikit-learn · AMRFinderPlus · Streamlit |
Project management where every person works through their own AI agent. Agents pick up unlocked tasks over MCP and report what they did and what it cost; humans approve the work.Hackathon prototype · TypeScript · Express · MCP · SQLite · React |
An offline advisor for flood-hit rice farmers in Bangladesh. A 6.1 MB on-device classifier answers only when it's confident, and otherwise says "not sure, ask your agriculture officer". In Bangla.Hack-Nation 2026 · World Bank challenge · PyTorch · ONNX · FastAPI · Next.js |
A layer coach: how many layers to wear today, picked from your own wardrobe, using your commute forecast, 10 years of local climate and your "too cold / too warm" feedback. Live.Next.js 15 · TypeScript · Postgres · Open-Meteo · Claude |
A marketplace where AI agents compete for your task and get paid in sats over Lightning. This repo is the web app; the full project is px404/EnrichedUranium.Team project · React · TypeScript · Tailwind · shadcn/ui |
A convolutional image classifier written from scratch in PyTorch: four conv blocks with BatchNorm, four augmentation pipelines, 20 classes, about 1.8M parameters.Python · PyTorch · torchvision · NumPy |
Also built with a team: Agent on a Leash, a wallet control layer that approves, declines or asks before an AI agent spends money (Swiss {ai} Weeks 2026, 15 hours).
📷 Car photography, ideally in the rain. Same thing I like in code: patience, light, and attention to detail.
Every animation on this page and in my repos is code, not a screen recording: each frame is drawn on a canvas, rendered in headless Chromium and encoded with ffmpeg. The source is in motion/.









