P.01 · Master's Thesis · TU Graz · Defence March 2026
LearnApp — documents to courses, with evidence.
LLM Pipeline · Semantic Chunking · LO Induction · Evaluation
The problem
Educational PDFs are unstructured walls of text. Turning them into courses means finding where topics actually begin and end, deciding what a learner should take away, and writing the material — work that doesn't scale by hand.
The system
A four-stage pipeline behind a React + FastAPI app: pre-processing (Docling with PyMuPDF fallback, pdfminer heuristics) turns PDF bytes into structured paragraphs; hybrid semantic chunking combines heuristic pre-chunking with GPT-4 boundary prediction and merge/split post-processing; LO induction groups chunks into learning objectives with a dual-provider design — LLM grouping vs. a seeded k-means baseline — so the approaches can be compared head-to-head; generation writes lessons and quizzes per objective, with free-text answers graded against reference content.
How it was evaluated
Instead of eyeballing outputs: a synthetic benchmark generates multi-topic PDFs with known gold boundaries, then measures precision, recall and F1 across six controlled experiments (document length, chunk count, section variability, topic similarity, formatting noise, post-processing) — 3,435 automated runs over 139K+ pages.
What came out
- LLM chunking beats the TextTiling-style heuristic on boundary F1; over-segmentation grows with document length — measured, not assumed.
- LLM-based LO grouping: +20% silhouette score and +14% topic similarity over the seeded k-means baseline.
- Expert review: chunk coherence 4.0/5, LO relevance 4.5/5, UI clarity 4.5/5 — and quiz quality 2.5/5, an honest weak spot that shaped future work.
P.02 · Product + Study · IntRS'23 Workshop, Singapore
Meal Architect — recommendation as an experiment.
LLM Generation · DALL·E · A/B User Study
A working meal-recommendation web app: enter your body stats and taste, get a recipe built to fit. Under the hood it doubles as a study — every user is silently assigned either our LLM pipeline or a conventional recipe-database baseline, and rates what they get blind. That comparison became a paper at the IntRS'23 workshop (Plangger, Rainer, Felfernig).
The pipeline
Biometrics (age, weight, height, gender, activity) feed a real nutrition engine — BMR via the Harris-Benedict equations, scaled to TDEE by activity factor (×1.2 to ×1.9), then split into macros (45% carbs / 35% fat / 20% protein). Those numbers, plus liked/disliked ingredients and meal type, become a structured prompt.
Biometrics→BMR · TDEE · macros→Prompt→50/50 split
The comparison — baseline vs. our pipeline
To find out whether an LLM actually helps, the app runs two recommenders behind one interface and assigns each user a random one (50/50 on page load), so ratings can be compared blind.
Baseline
Spoonacular database
A conventional recommender: query a fixed database of thousands of existing recipes and return the closest match to the user's calories and ingredients. Reliable, but bounded by what already exists — and its stock photos rarely match the exact dish.
fixed recipe setnearest-matchstock photos
Meal Architect · our pipeline
LLM-generated, made to fit
Our own pipeline builds a structured prompt from the user's nutrition targets and taste, has GPT compose a brand-new recipe around exactly those constraints, returns it as strict JSON (the model was primed on valid examples so parsing stays reliable), then paints a matching dish with DALL·E 3 — because an invented meal has no photo to borrow.
unbounded recipesfits your macros + likesAI-generated photo
What the study found
Users rated the meals they were served 0–5 across four axes, not knowing which recommender produced them. Our pipeline won every category:
Picture3.583.97
Would eat3.894.31
Ingredients3.634.29
Satisfaction3.824.00
■ Baseline (Spoonacular DB)■ Meal Architect (our pipeline)
Biggest gap was ingredients (+18%) — the LLM's freedom to build around exactly what a user likes, rather than match the closest thing in a database, is the whole thesis in one bar.
Honest limits
DALL·E images were slow and stylistically inconsistent, and the ingredient picker needed a lot of scrolling on mobile. The paper lays out the roadmap it opened: weekly meal plans with shopping lists, cook-from-your-fridge photo input, and a conversational ordering mode.
P.03 · Computer Vision
Tennis Pose Analysis — a coach in the pipeline.
Pose Estimation · Biomechanics
What it does
Upload a serve video; the pipeline runs pose estimation, detects the five serve phases (start, loading, acceleration, contact, finish), checks biomechanical markers like elbow angle at contact, and renders an annotated video plus structured feedback. Built as backend services on PyTorch and Flask, deployed on AWS — a background in competitive tennis meant the checks weren't invented at a desk.
P.04 · Simulation / Graphics
2D Physics Engine — the fundamentals, from scratch.
Rigid Bodies · SAT · Fracture
What's inside
No physics library: custom rigid-body dynamics with separating-axis collision detection, a mass-spring bridge, Voronoi fracture, motion blur, and runtime debug controls. It's the non-AI counterweight in the portfolio — proof the math and the rendering loop hold up without a framework underneath. The tile on the main page runs a miniature of it live; the repo has the full engine.