Samuel Konkoľ

I’m an Applied Informatics student who treats software architecture as a system to be profiled, debugged, and optimized. I focus on high-performance Python and C++ work, especially where backend infrastructure and local AI pipelines meet.

In my Local NLP Fact-Checking Engine, I built a fully local retrieval pipeline that runs offline on consumer hardware (RTX 4060) with zero paid APIs—ASR, embeddings, and matching included. I engineered a hybrid retrieval ladder combining lexical RapidFuzz matching with dense BGE-M3 embeddings, reaching 0.994 Success@1 on a hand-built Slovak gold set.

I also focus on reliability and “know when to abstain.” I trained a learned verifier (logistic regression over match signals) scoring 0.983 open-world abstention F1, so the system can abstain when no prior fact-check exists. To convert messy debate audio into matchable statements, I integrated WhisperX ASR.

I’m also developing practical systems like a cross-platform YouTube bot (asynchronous, OAuth 2.0 secured, with quota-resilient exponential backoff) and a Smart Recipe Extractor browser extension that structures recipes and converts every measurement to metric units, adding scannable QR codes for ingredients.