CV

My Professional CV

Contact Information

Name Masoud Rezvani
Professional Title ML Engineer & Data Scientist
Email s.masoudrezvani@gmail.com
Location Amsterdam,

Professional Summary

Machine learning engineer with 5+ years in data, including ~3 years deploying and operating production ML systems: fraud-detection models scoring ~200K trips/day ($500K/year prevented losses) and forecasting/pricing pipelines across 10 cities with automated drift monitoring and retraining (Airflow, Spark). Completed an MSc in Data Science at the University of Amsterdam while building production retrieval and LLM systems: hybrid search, reranking, fine-tuning (QLoRA), and evaluation. Published in Expert Systems with Applications (ESWA, 2025).

Experience

  • 2026 - Present

    AI Engineer Intern
    Talk360
    International calling service (Amsterdam).
    • Built a production information-retrieval system: hybrid lexical (BM25) + semantic (ChromaDB) retrieval with cross-encoder reranking; 98% faithfulness on an internal evaluation set, with inline citations and prompt-injection defenses.
    • Built Gemini-based agents with n8n ETL pipelines ingesting Mixpanel usage logs to stream real-time Slack alerts for free-minute abuse, credit-sharing fraud, and low call ratings, cutting incident detection from hours to less than a minute and preventing support ticket surges.
    • Built a semantic caching layer to reduce repeat-query API costs; implemented an LLM-as-judge regression pipeline (GPT-4o) for pre-deploy validation.
  • 2023 - 2025

    Data Scientist
    Baly.iq (Rocket Internet)
    The #1 super app in Iraq.
    • Built, tuned, and deployed XGBoost and logistic-regression fraud-detection models scoring ~200K trips/day of live ride-hailing traffic, preventing an estimated $500K/year in fraud losses at a 0.8% false-positive rate without creating user friction.
    • Built and deployed surge-pricing models across 10 major cities, reducing rider wait times by 5% (from 8.5 min to 8.0 min) during peak morning hours in key markets.
    • Designed and analyzed switchback A/B tests for localized surge-pricing algorithms, informing the full rollout of dynamic fare multipliers that lifted active driver supply by 12% and increased request fulfillment rate by 7% during peak hours.
    • Partnered with engineering to productionize ML models via APIs and batch pipelines (Airflow, Spark) with automated drift monitoring and retraining; operated production monitoring (Grafana, Netdata) on AWS.
  • 2020 - 2023

    Data Analyst
    Snapp! (Rocket Internet)
    The #1 ride-hailing super app in Iran.
    • Owned dashboards for trips, cancellations, and fraud, serving stakeholders across the organization.
    • Wrote complex SQL queries against large-scale trip and event data to answer ad-hoc business questions on rider behavior, pricing, and fraud.
    • Automated recurring reports with Python and scheduled queries, saving the team ~10 hours per week.

Education

  • 2025 - 2026

    Master of Science
    University of Amsterdam (UvA)
    Data Science
    • Thesis: Evaluating the Muon Optimizer — From Symbolic Regression Geometry to High-Dimensional LLM Adaptation
  • 2017 - 2019

    MBA
    University of Tehran
    Business Administration
  • 2012 - 2016

    BSc
    Amirkabir University of Technology (Tehran Polytechnic)
    Industrial Engineering

Publications

Skills

Languages: Python, SQL, PySpark
ML Frameworks: PyTorch, HuggingFace (Transformers, PEFT), LightGBM, XGBoost, scikit-learn, Learning-to-Rank & Information Retrieval
NLP & LLMs: RAG, Reranking, LLM-as-judge evaluation, Fine-tuning (QLoRA/PEFT), LangChain
ML Infrastructure: MLflow, Weights & Biases, Docker, Airflow, FastAPI, Git, CI/CD, Grafana, Netdata
Cloud & Databases: AWS, FAISS, ChromaDB, PostgreSQL, MySQL, ClickHouse

Languages

Persian : Native speaker
English : Fluent
Dutch : Beginner

Interests

Research Interests: Large Language Models, Deep Learning, Generative AI, Graph Neural Networks, MLOps

Certificates

  • AI Engineer Core Track (LLM Engineering, RAG, QLoRA, Agents) - Udemy (2026)
  • Machine Learning System Design - ByteByteGo (2026)
  • Stanford CME295 — Transformers and Large Language Models - Stanford University (2026)
  • Stanford CS336 — Language Modeling from Scratch - Stanford University (2025)
  • Associate DevOps Engineer - Arvancloud (2025)
  • Graph Search, Shortest Paths, and Data Structures (Algorithms Specialization, Part 2) - Coursera (2024)
  • Applied Statistical Modeling for Data Analysis in R - Udemy (2023)

Projects

  • Learning-to-Rank News Recommendation System
    • Corrected position bias in click logs via inverse-propensity weighting, recovering ~6% NDCG@10 (0.745 to 0.802) over naive click training in a controlled synthetic experiment.
    • Hybrid retrieval stack: BM25 lexical + FAISS approximate nearest-neighbour candidate generation feeding a LambdaMART learning-to-rank model.
  • LLM Support-Ticket Triage
    • Diagnosed overfitting from per-epoch eval curves and ran a Weights & Biases LoRA rank ablation (r = 4-32).
    • Pydantic-validated data pipeline with a pytest suite for reproducible training and evaluation.
  • Vision-Language Model for DocVQA
    • Fused visual embeddings (ResNet + CLIP) with OCR text outputs fed into a transformer LLM.
    • Evaluated on standard DocVQA benchmarks; trained end-to-end in PyTorch.
  • SemArt: Texture Bias in CNNs on Historical Art
    • Benchmarked ResNet-50 and CLIP using InfoNCE contrastive learning.
    • Quantified texture bias via UMAP visualization and inter-class distance metrics.
  • Interactive Model Interpretation Dashboard
    • Backend built with FastAPI; frontend exposes real-time SHAP dependency plots.
    • Fully containerized with Docker for reproducible deployment.
  • Algorithmic Trading System
    • Feature engineering over price-action patterns (support/resistance, volume profiles, candlestick structures) with Polars.
    • All experiments tracked with Weights & Biases for systematic ablations across feature sets and models.
  • LLM Arabic Fraud Detection
    • Engineered data-cleaning and tokenization strategies suited for Arabic morphology.
    • Built strict evaluation pipelines for reproducible, production-ready results.
  • CARE-GNN Reconstruction
    • Full reconstruction of heterogeneous graph construction, label-aware similarity estimation, and RL-based neighbour selector.
    • Surfaced underspecified implementation details around camouflage-aware aggregation and the RL exploration-exploitation tradeoff.