M.Sc. Information Systems at the Karlsruhe Institute of Technology. I work on measuring things about machine learning systems that are usually only described — how much a language model's answer really depends on how you phrased the question, whether an agent's numbers are actually grounded in its evidence, what a strategy actually returned.

• Prompt sensitivity in LLMs — developing Functional Information (FI_in), an information-theoretic measure of how much a model's answer changes when a question is paraphrased but its meaning is not. Ongoing research at KIT. → PromptSensitivityFI
• DrawingIQ — turning an archive of ~11,000 engineering drawings into something you can ask questions of. A vision LLM and a deterministic parser extract a structured record from each sheet, and a tool-using agent answers over the result while citing the drawings behind every claim. Extraction quality is scored against a gold set rather than asserted. → drawing-intelligence
• Grounded agentic reasoning — an agent over automotive crash telemetry in which every quantitative claim must cite the tool result it came from, enforced by an automatic citation checker and measured by an evaluation harness. → DriveTSLM
• M.Sc. thesis, expected March 2027.

Alongside this I work as a freelance data scientist at Dark Sky Data LLC, building analytics pipelines for automotive finance and warranty clients.

Automated Market Makers: Toward More Profitable Liquidity Provisioning Strategies Thanos Drossos, Daniel Kirste, Niclas Kannengießer, Ali Sunyaev SAC '25 — 40th ACM/SIGAPP Symposium on Applied Computing, Catania, Italy

DOI · arXiv · code & data

The paper measures the realised return of 30,058 closed Uniswap v3 liquidity positions across 9 pools over roughly 700 days, decomposing each outcome into impermanent loss and collected fees. The repository reproduces every figure in the paper from committed data.

• A wearable that detects nocturnal jaw clenching — three electrodes across the forehead pick up the temporalis muscle, a per-user linear SVM classifies the frequency spectrum on an Arduino inside the sampling budget of a 1 kHz signal, and escalating feedback interrupts the episode before the teeth take damage. The model is trained per wearer because EMG amplitude varies far too much between people for a shared one to work. KIT Ubiquitous Computing, WS 2025/26. → JawClenchingProject

Empirical evaluation of LLMs · robustness and prompt sensitivity · agentic systems grounded in real data · multimodal document understanding · on-device machine learning · reproducible research

thanosdrossos.github.io · Karlsruhe, Germany