AI/ML engineer building production machine-learning systems, data pipelines, and network-automation platforms. I work at the intersection of applied ML and real infrastructure — from GNNs and transformer fine-tuning to Airflow pipelines and LLM agents.

• 🏢 Junior Data & AI Engineer at Makedonski Telekom AD
• 🔬 Machine Learning Researcher at the Macedonian Academy of Sciences & Arts (MASA)
• 📄 Co-author of two research papers — ADMET property prediction (GNNs) and formal-language ambiguity analysis
• 🎓 B.Sc. Computer Science & Engineering at Ss. Cyril & Methodius University (FCSE) — finishing my degree, only the thesis remains
• 🌱 Focused on GNNs, RAG, LLM agents, and MLOps
• 📫 Reach me at martin.stamenov03@gmail.com · Skopje, North Macedonia

Focus areas: GNNs · Transformer fine-tuning · Reinforcement learning · RAG & LLM agents · ETL & data pipelines · Network automation (NETCONF, SNMP)

🏢 Junior Data & AI Engineer — Makedonski Telekom AD · Sep 2025 – Present

• Building NetFlow, an in-house network-automation platform (~90 service modules, 530+ API endpoints) that turns manual multi-vendor router changes into automated rollouts, with a workflow engine, scheduler, approval gates, and rollback — now adding an LLM agent layer.
• Added ML traffic and capacity forecasting to Zabbix network monitoring using scikit-learn and statsmodels inside scheduled Airflow pipelines.
• Built an AI assistant for the Ministry of Digitalization (a CKAN extension over data.gov.mk) on the OpenAI Assistants API, answering natural-language questions over public datasets in Macedonian.

🏢 Data Engineer Intern — Makedonski Telekom AD · Mar 2025 – Sep 2025

• Designed Apache Airflow pipelines pulling data from Huawei NCE, SNMP, and optical telemetry into a central PostgreSQL database.
• Built Grafana dashboards for DWDM optical network monitoring with automated alerts.
• Developed an internal RAG chatbot for technical support with OpenAI embeddings and vector search over ChromaDB; fine-tuned Hugging Face models on telecom data.

🔬 Machine Learning Researcher (Intern) — Macedonian Academy of Sciences & Arts (MASA) · Aug 2025 – Present

• Benchmarked GNNs against molecular foundation models (MolCLR, ChemBERTa) for ADMET property prediction across six TDC datasets under scaffold-split (~2,400 training runs).
• Standardized datasets, engineered molecular features with RDKit, and compared bio-inspired hyperparameter optimization methods (PSO, GA, ABC, TPE) using AUC/ROC.

FastAPI · PostgreSQL/pgvector · Next.js · Playwright An agent that turns one product idea into channel-specific posts across six platforms. A 7-stage pipeline (research → strategy → copy → self-review → repair) drafts each campaign; a human approves it over Telegram, and a rehearsal mode tests real platform calls before publishing.

Flask · PostgreSQL · NETCONF · Kubernetes In-house, NSO-style network provisioning platform (~90 service modules, 530+ API endpoints) for multi-vendor routers, with a workflow engine, job scheduler, approval gates, and automatic rollback. Currently adding an LLM agent layer.

PyTorch Geometric · GraphSAGE · Flask · React Full-stack event recommender for North Macedonia: a user-event graph with GraphSAGE link prediction, served through Flask and React, over data scraped from six local sources.

XGBoost · LightGBM · MLflow · FastAPI · Docker · Grafana End-to-end MLOps pipeline for the IEEE-CIS fraud dataset: Optuna-tuned LightGBM/XGBoost (AUC 0.85 → 0.94) plus a real-time streaming scorer with three-tier ALLOW/REVIEW/BLOCK routing. Packaged as an 8-service Docker Compose stack with MLflow tracking, a FastAPI endpoint (~146 ms), and Prometheus/Grafana monitoring.

ADMET-Bench: Task-Dependent Performance of GNNs and Pretrained Models in ADMET Prediction Co-first author — in preparation for IEEE JBHI. Benchmarks five model families across six TDC endpoints under scaffold-split (~2,400 training runs); shows ADMET predictability is endpoint-specific.

Bounded-Depth Ambiguity Detection in Context-Free Grammars First author — submitted to ICT Innovations 2026 (Springer). Bounded search over a 50-grammar benchmark that finds ambiguity witnesses in CFGs, cross-validated by an independent SAT encoding (100% agreement).

B.Sc. Computer Science & Engineering — Ss. Cyril & Methodius University (FCSE), Skopje · 2022 – 2026 Relevant coursework: Machine Learning, Artificial Intelligence (LLMs, NLP, RL, Agents).

🎯 Coursework complete — only the diploma thesis remains before graduation.

Languages: Macedonian (Native) · English (Professional)

Always open to learning, collaborating, or discussing interesting AI ideas!