Masoud Rezvani

I am Masoud Rezvani (Masoud Rezvaninejad), an ML Engineer and Data Scientist currently completing my MSc in Data Science at the University of Amsterdam (UvA). My focus is on building production-grade ML systems — reading a paper, turning it into working code, and shipping it.

I am currently an AI Automation Intern at Talk360, where I built an HR policy assistant using hybrid RAG (ChromaDB semantic search + BM25 keyword retrieval) reranked by a Cross-Encoder, achieving 98% faithfulness on a 22-question golden dataset. I also designed a semantic caching layer to cut repeat-query API costs, an LLM-as-judge regression pipeline (GPT-4o) for pre-deploy validation, and autonomous LangChain + Gemini agents that interpret execution logs and diagnose system failures.

Previously, as a Data Scientist at Baly.iq (Rocket Internet), I built XGBoost and logistic regression fraud-detection models reaching 95% accuracy, contributing to an estimated $500K in annual savings. I deployed containerised fraud-detection services with Docker and Ansible across multi-server environments, maintained AWS EC2 infrastructure (Grafana, Airflow, Netdata), and improved MySQL/PostgreSQL query performance by 20% using Percona PMM. At Snapp! before that, I applied CNNs and GNNs to detect fraud patterns and fraud rings, improving detection efficiency by 10%, and built the Airflow/Docker pipeline infrastructure that later went into production at Baly.iq.

My MSc thesis investigates the Muon optimizer — evaluating its geometry on symbolic regression loss surfaces and its behaviour in high-dimensional LLM fine-tuning via LoRA. I have a published paper in Expert Systems with Applications (ESWA, October 2025): WDAE-GAN, a hybrid dual-autoencoder GAN with wavelet denoising for credit card fraud detection (DOI: 10.1016/j.eswa.2025.130078).

Technical stack: Python · PyTorch · HuggingFace · LangChain · MLflow · FastAPI · Docker · Airflow · AWS · ChromaDB · PostgreSQL · Rust (intermediate)