AI / ML Engineer — LLMs · RAG · agentic systems
I build production LLM systems — agentic pipelines, retrieval and multilingual NLP. Currently Team Lead / AI Engineer at Factum AI, shipping a multilingual legal chatbot across 15+ European markets.
// my work — drag to spin it, or click a node to jump to the details below
Lead engineering for a multilingual multi-agent legal chatbot — live in Macedonia & Serbia, with 15+ markets operational. I own technical direction across agentic flows, data ingestion, web scraping, Celery async and the relational/non-relational database layer, and mentor 7 junior engineers.
factum.lawBackend for a decentralized GPU compute marketplace (CTO: Prof. Panche Ribarski). Designed an async resource-reallocation algorithm scheduling LLM workloads across a distributed machine pool, served open models via vLLM in per-model Docker images, built the frontend, and wired Solidity smart contracts + the native token into payments and metering.
gptprotocol.orgBuilt a chatbot for Alzheimer’s researchers powered by hybrid RAG (BM25 + transformer-based dense retrieval) over tens of thousands of peer-reviewed papers and a Neo4j knowledge graph — part of MANU’s diagnosis-support project under Prof. Igor Mishkovski.
manu.edu.mk“On the Capabilities of LLMs for Style Detection and Lyric Generation in the Macedonian Language.” A custom dataset of 36 authors and 1,199 songs, 18 synthetic datasets across style-transfer techniques — 80% F1 on authorship over generated text, 70% over real poetry. Supervised by Prof. Sonja Gievska & Prof. Martina Toshevska; later guest-lectured the FCSE NLP course on it.
Member of Team Polaris — developed AI agents that outperformed the field in the multi-agent challenge at RoboMak 2023 (IEEE Macedonia), taking 1st place in the AI Agents category.
read the writeupPublished Macedonian-language poet, with readings at literary festivals across Macedonia. Half-seriously, the reason I got into language models in the first place — and the motivation behind my thesis.
I’m an AI/ML engineer based in Macedonia. Most of my work lives where language models meet real systems — agentic flows, retrieval (BM25 + dense), knowledge graphs, and the unglamorous backend that keeps them running in production.
Before Factum I built backend for a decentralized GPU compute marketplace and a hybrid-RAG research assistant at MANU. I also write Macedonian poetry — which, half-seriously, is what pulled me into language models: my BSc thesis was on LLMs for style detection and poetry generation in low-resource Macedonian.
// basalt columns — move your mouse across to raise them