MSc Informatics at the University of Zurich. Artificial Intelligence major, Data Science minor. Based in Luzern, Switzerland.
I work on making model outputs checkable: whether an answer can be traced to a source, and whether a system fails evenly across the people it is used on. Before Zurich I did a BSc in Computer Science at Royal Holloway, University of London.
Currently looking for a working-student or research-assistant role in Switzerland, 15 h/week during term and full-time in the lecture-free period. Graduating July 2027.
Recursive leverage in yield-bearing stablecoins · full report (PDF)
MSc project with Dawid Rymarczyk, supervised by Prof. Claudio Tessone and Krzysztof Gogol at UZH. Large wallets loop yield-bearing stablecoins on Aave v3 and Morpho Blue to stack extra yield. We ask whether they are really following a handful of shared strategies, how those strategies differ in risk management, and whether the on-chain state of a position can tell you a stressful day is coming. HDBSCAN/UMAP for the cohorts, XGBoost with SHAP attribution for the prediction, ROC-AUC 0.90 and 0.98.
Multi-token DeFi staking pool on Base · overview
Most staking apps pay you by printing a new token. This one does not: you deposit WETH, USDC or EURC, the contract routes it into a Morpho MetaMorpho vault, and a daily Chainlink Automation harvest returns the yield in the same token you put in. Team project for the Seminar in Blockchain Programming at UZH, graded 5.5/6.
Detecting machine-generated academic essays with a multi-view ensemble (DeBERTa-v3, RoBERTa likelihood features, stylometry). The interesting result is not the F1 score. The evaluation asks whether false positives land disproportionately on non-native English writers, because a detector with strong headline accuracy can still be unusable for anyone writing in a second language. Four-person team; I built the preprocessing and tokenisation over an 800k-sample corpus and the non-native hold-out set.
Visual odometry from a phone camera · full report (PDF)
Monocular VO (FAST, KLT, PnP RANSAC) with 100% pose success across 13,000+ frames. A single camera recovers its path but not its scale, and a fast rotation is enough to lose the track, so we filmed a Swiss cablecar on a Samsung S20 and used the phone's gyroscope to survive the rotations and its IMU to put the trajectory into metres: 4.2x scale recovery, matching barometric elevation to within 1.2%.
Knowledge-graph QA chatbot · report (PDF)
Answers questions from a Wikidata-derived knowledge graph. SPARQL first, embeddings when the graph has no hit, and only then a local LLM on Ollama, so that an answer points at a source instead of being generated freely.
Built during a data science internship at Zappi in London: a Python pipeline using XGBoost and SHAP to identify which features of a product concept actually move outcome metrics.
- Languages: Python · Java · TypeScript/JavaScript · Solidity · SQL · C++ · Bash · LaTeX
- ML: PyTorch · scikit-learn · XGBoost · SHAP · HDBSCAN/UMAP · SpaCy · transformers · Ollama
- Data and web: pandas · NumPy · Jupyter · Next.js · React · Prisma · tRPC · MySQL
- Vision: OpenCV · visual odometry · camera and IMU sensor fusion
- Chain: Hardhat · ethers.js · OpenZeppelin · Chainlink · Gnosis Safe
- Tools: Git · Linux (Arch/Manjaro, Debian) · Docker · Vim
iglikristo.xyz · LinkedIn · igli.kristo@uzh.ch
Languages: English (native), Albanian (fluent), German (A1.2, continuing each semester at the UZH Language Center; A2.2 expected by graduation).