Hi, I'm Andreas

MSc Computer Science student at Aalborg University · Machine learning research · LLMs · Recommender systems

I'm a master's student in computer science at Aalborg University, Denmark, working on machine learning research. I care about the gap between what machine learning methods are claimed to do and what they actually do, so I analyse methods from first principles, reproduce published results under controlled conditions, and build small, well-tested research software that makes results easy to verify.

• Machine learning research: how training choices such as loss functions, reweighting and regularisation change what a model actually learns, backed by proofs and controlled experiments.
• LLMs and generative AI: parameter-efficient fine-tuning with LoRA and QLoRA, 4-bit quantisation, and open models such as Qwen.
• LLM inference optimisation: distilling large language models into smaller, faster ones, the topic of my bachelor's project.
• Recommender systems: multimodal and cold-start recommendation, graph-based recommenders, and the reproducibility of published models.
• Reproducible research: controlled experiments, sound statistics, and long training runs on GPU clusters.

Distil an LLM classification prompt into a small, fast model that runs on a CPU. Describe the task in a short YAML file and give it unlabelled texts: an instruction-tuned LLM labels them with soft targets, a small encoder learns to imitate them, and the result is exported to ONNX, optionally int8. It continues my bachelor's project on inference optimisation by distilling LLMs.

Try the live demo · Documentation · Code · pip install tinydistill

flowchart LR
    A["Task prompt +<br/>unlabelled texts"] --> B["LLM teacher(s)<br/>Qwen · Mistral · Phi · ..."]
    T["tune: check the<br/>task file"] -.-> B
    S["Synthetic texts<br/>(synth)"] -.-> B
    V["vLLM server ·<br/>hosted API"] -.-> B
    K["A few gold labels<br/>per class"] -.-> B
    K -.-> C
    B -- "soft labels<br/>(ensembled)" --> C["Small encoder<br/>student"]
    C --> D["ONNX / int8<br/>export"]
    D --> E["Fast CPU<br/>inference"]
    D --> W["Browser · server · Hub ·<br/>Python (Distiller)"]
    E -.-> F["Cascade: unsure<br/>texts to the LLM"]

Results on Banking77 (77 intents, 3,080 held-out test texts; latency per text on an RTX 4090 for the teacher and 2 CPU threads for the students; the students learn from the teacher's soft labels plus 8 gold labels per class; all benchmarks):

The students are more accurate than their teacher and hundreds of times faster on a CPU than the teacher on a GPU.

An interactive, in-browser demo of knowledge distillation. A teacher network and two small students train side by side on 2D data, one from the labels alone and one from the teacher's soft targets; sliders for the temperature, the mixing weight and the student size show when distillation helps. Written from scratch in TypeScript, with no ML library and gradient-checked backpropagation. Try the live demo · Code

Mathematical bugs I found and reported in widely used ML libraries, each with a minimal reproduction, and fixed in a pull request where I could.

Languages

Machine learning

Methods and models

LoRA QLoRA PEFT Quantisation Knowledge distillation Inference optimisation Hashing Qwen Transformers Recommender systems Graph neural networks Multimodal learning Class imbalance Statistical testing Closed-form recommenders Cold-start recommendation Pre-registered evaluation

Research and infrastructure

Web (earlier projects)

• MSc in Computer Science, Aalborg University (ongoing)
• BSc in Computer Science, Aalborg University, with a bachelor's project on inference optimisation by distilling LLMs

• Writing up research on machine learning training objectives
• Pre-thesis project on reproducibility and cold-start ranking in multimodal recommender systems
• Auditing ML libraries for mathematical bugs (contributions)
• Open to collaborations on machine learning research and tooling