Stefan Decker

Built Order Robot as a solo developer from concept to production—an AI-powered tool that automates the conversion of PDF and email-based B2B orders into Shopify Draft Orders.

THE PROBLEM
D2C brands receiving wholesale orders via PDF or email manually re-enter them into Shopify, a tedious and error-prone process. Order Robot eliminates this by using a large language model (LLM)-based document parsing to extract line items, quantities, pricing, and customer data from unstructured order documents, then automatically creates Draft Orders via the Shopify GraphQL API.

TECHNICAL HIGHLIGHTS
LLM-powered document parsing with structured output extraction, Shopify GraphQL API integration (Draft Orders, B2B pricing with Company/Location/Catalog/Price List hierarchy), email-based order intake, and a web-based interface for order review and confirmation.

This project demonstrates end-to-end AI product development: identifying a real business pain point, designing the solution architecture, building the full stack, and deploying to production all as a single developer.

Built SpinForge (Zwift AI Workout Generator) as a solo project, a web application that generates custom indoor cycling workouts for the Zwift platform using AI.

Users specify their training preferences (duration, focus area, intensity) and receive a structured workout tailored to their goals. The key differentiator: workouts can be refined through natural language conversation. Users describe changes in plain English (e.g., 'make the intervals shorter but more intense' or 'add a longer warm-up'), and the LLM modifies the workout accordingly while ensuring it still meets the original constraints.

TECHNICAL IMPLEMENTATION
Large language model (LLM)-powered workout generation with constraint validation (ensuring duration, intensity zones, and training focus remain consistent), conversational refinement interface for iterative workout editing, structured output generation in Zwift-compatible workout format, and deployment on Fly.io.

This project showcases practical LLM application design: not just generating content, but maintaining structured constraints while allowing flexible natural language interaction, a pattern applicable to any domain where users need AI-assisted creation within defined parameters.

I built a German-language question-answering system that lets users ask natural language questions to politicians and receive relevant answers from their social media statements. The system uses a Dense Passage Retriever for semantic document retrieval and a fine-tuned Electra transformer model as a ranker to identify the most relevant answers, even when they don't share keywords with the question.

• Created custom training and evaluation datasets by hand-labeling 144 questions across 525 political tweets.
• Fine-tuned both retriever and ranker on this domain-specific data, significantly boosting recall and F1 scores.
• Deployed as a Flask web app on Google Cloud Run with a PostgreSQL-backed FAISS document store.
• Built the full pipeline solo: data collection via the Twitter API, text normalization, model training, evaluation, and cloud deployment.

The Deep House Spy is an artist classification project based on a convolutional neural network (CNN). CNNs have proven to be really effective for image classification and can also work with audio when put in the right form.

The idea for the project came to me while listening to a deep house DJ set on SoundCloud. I really liked the song that I was listening to, but Shazam just wouldn't recognize it. This happens very often, and the reason is (most of the time) that Shazam doesn't have the song in its database because it hasn't been released yet.

My approach to solving this problem was to learn the styles of artists by using already released songs and then identifying the respective artists of unreleased songs.