Portfolio
MSc Thesis
Combined classification and regression models with traffic-aware thresholds to reduce power consumption while maintaining performance. Included explainability using SHAP values.
BSc Thesis
Making board games more accessible for people with visual impairment via camera + image recognition, converting physical cards into a digital hand. Tested on Catan with high accuracy.
LLM Application
Skincare web app with an integrated LLM assistant running Llama 3 locally through Ollama, with simple memory functionality and a REST API connecting three components.
Meta-Learning
Extends WACO with meta-learning to adapt across hardware with minimal data. On 300 samples from a new machine, our meta-trained model outperformed the original WACO.
Computer Vision
Transfer learning with ConvNeXT for dermoscopy images. Best model reached 0.88 accuracy on the test set.
RAG
RAG pipeline using Pinecone vector DB. Best config: e5-base + Mistral, beating GPT-3.5-turbo on this dataset.
RAG
Financial QA using ARKMan, hybrid search, chain-of-thought, and query decomposition on the FinQA dataset.
NLP
3000 docs from 20 Newsgroups. Best coherence with 10 topics and α = β = 0.1.
Computer Vision
CNN for 12 species from bark images. Ablation study highlights: residuals, normalization, batch norm. Added OOD vs ID task (100% on that split).
Optimization
11-node grid model minimizing generation cost while meeting demand. Best found cost: 186.29 SEK.
Fine-tuning
Comparing LoRA with feature extraction for task adaptation. LoRA achieved Acc/F1 ≈ 0.93 vs 0.81.
NLP
Best baseline: Logistic Regression + TF-IDF (Acc 0.89) on annotated social media comments.
Reinforcement Learning
MCTS with UCT. Win rate: 0.86 vs random; 0.83 vs blocking strategy.
Recommendation
Hybrid preferences using prior ratings, genres, and similar users.
NLP
Deterministic bigram per category achieved highest semantic similarity (0.56) and strong BLEU on this dataset.
NLP
IBM Model 1 on parallel EN-SV-DE-FR corpus.