NLP Researcher | MSc Information Technology, Metropolia University of Applied Sciences — Graduated May 2026 (GPA 5.00/5.00) | Based in Espoo, Finland
Applying for PhD positions in NLP/AI, with a research focus on multilingual NLP, clinical/medical NLP, LLM evaluation, and trustworthy AI — specifically hallucination detection, grounding, and reliability of LLM-generated text in high-stakes domains.
Solo, independent submission — zero-shot LLM-as-judge hallucination span detection for vision-language model outputs. Ranked 13/28 teams (Cor+Lbl) among international participants. System uses Claude Haiku 4.5 as primary judge with Claude Sonnet 5 for second-opinion verification on ambiguous error categories, plus rule-based post-processing. 📄 System description paper submitted to OpenReview; arXiv preprint in progress. 🔗 shroom-visions-hallucination-detection
Extending the SHROOM-Visions hallucination-detection methodology to medical/clinical text, built specifically for PhD research-plan development. Combines direct source-fidelity checking with retrieval-augmented grounding (sentence-transformer embeddings + FAISS) against trusted medical references (MedlinePlus), verifying LLM-generated clinical summaries for invented claims, mischaracterizations, and unsupported medical assertions. 🔗 Medical-grounding-verification
Fine-tuned FinBERT (TurkuNLP/bert-base-finnish-cased-v1) on labeled Finnish news data, benchmarked against a TF-IDF + Logistic Regression baseline. FinBERT achieved 0.787 macro F1 vs. 0.745 baseline. Includes class-imbalance and data-scarcity error analysis. 🔗 finnish-news-topic-classification
Context-Aware Document-Level Location Selection for Finnish News Articles: A Hybrid Rule-Based and LLM Approach Developed in collaboration with Superhood Oy — a Finnish neighbourhood-level news platform. Awarded grade 5/5 — the highest grade at Metropolia University of Applied Sciences.
Built and evaluated a four-configuration NLP pipeline for extracting and ranking geographic locations at postal-code granularity from Finnish news articles, combining:
- Stanza for Finnish NER and morphological normalisation
- Geoapify for dynamic geocoding and candidate resolution
- Postal-first hierarchical ranking for geographic level disambiguation
- Model Context Protocol (MCP) with Llama 3.3 70B via Groq for contextual reasoning
Key result: 83.33% exact-match accuracy with the full hybrid configuration across 60 Finnish news articles — a 16.66-point improvement over the rule-based baseline. Key finding: Systematic lemmatization failures on Finnish postal-level names point to a data-coverage gap in Finnish NER training corpora, not a tool-architecture limitation.
📄 Published thesis: URN:NBN:fi:amk-2026051311846 📊 Evaluation dataset & results: thesis-nlp-evaluation
| Area | Tools & Technologies |
|---|---|
| LLMs & Evaluation | Claude API (Haiku, Sonnet), LLM-as-judge methodology, hallucination/error taxonomy design, prompt engineering, zero-shot evaluation |
| Retrieval & Grounding | RAG pipelines, sentence-transformers, FAISS (vector similarity search), embedding-based retrieval |
| NLP Modeling | FinBERT fine-tuning, transfer learning, Stanza, Finnish NER, lemmatization, TF-IDF, transformer architectures |
| LLM Integration & Agents | MCP (Model Context Protocol), FastMCP, Groq, Llama 3.3 70B |
| ML Foundations | Regression, classification, ablation study design, class-imbalance handling, error analysis |
| Backend | Python, FastAPI, Flask, REST APIs |
| Databases | PostgreSQL, MongoDB, MySQL |
| Data | Pandas, NumPy, evaluation framework design |
Hands-on exploration of Model Context Protocol — the tool-orchestration framework used in the thesis disambiguation layer.
Notebooks and experiments from NLP and Machine Learning specialisation coursework.
Prototype language-learning app providing real-time feedback on Finnish grammar and pronunciation.
Backend REST API development for a VoIP application. Implemented CRUD operations and database integration using Flask and MySQL.