Engin Deniz Dogu

New York, NY

LinkedIn · GitHub · Website

Summary

Data scientist with 3 years of experience in data science and applied machine learning, plus 2 years in enterprise risk consulting. Engineered fraud classification models that prevented $4M in losses. M.S. in Machine Learning (expected December 2026). Research experience in agentic workflows and semantic caching. Using AI-assisted coding to accelerate prototyping and development.

Education

• Master of Science in Machine Learning (GPA: 4.0/4.0)
• Mentored 45+ students in Mathematical Foundations of Machine Learning as a graduate course assistant

• Bachelor of Science in Computer Engineering & Bachelor of Arts in Economics

Skills & Certifications

• Data Science: Python, Jupyter, Pandas, NumPy, scikit-learn, PyTorch, Anomaly Detection, XGBoost, Random Forest
• AI: LLMs, Prompt Engineering, Hugging Face Transformers, Semantic Search (FAISS), AI-assisted Coding (OpenAI, Claude)
• Data Engineering: Statistical Analysis, SQL, PostgreSQL, ETL, PySpark, IBM DataStage, SAP BusinessObjects, Data Visualization
• MLOps: Docker, Kubernetes, MLflow, KNIME, Git
• Certifications: AWS Cloud Practitioner, AWS Partner Accreditation (Technical), Azure Data Fundamentals, ISO 27001 Lead Auditor

Experience

• Enhanced the fraud detection pipeline by training and tuning classification models using Python, KNIME, scikit-learn, and Optuna, improving precision from 62% to 70% while maintaining 3 second end-to-end inference, and preventing $4M in losses in FY2025.
• Containerized KNIME and Python environments using Docker and Kubernetes, standardizing production workflow deployments and achieving 99.9% uptime while removing integration bottlenecks for 12 data scientists.
• Designed KNIME dashboards tracking accuracy, precision, recall, and drift across 60+ production models, replacing reactive, business-unit reported issues with weekly automated reviews and threshold-based alerts that informed retraining decisions.

• Restored integrity of a 3,000+ record repair shop database (50% missing data) by building an automated Python preprocessing pipeline with Google Maps API, cutting manual data cleanup time by 30% and enabling accurate real-time tow truck dispatching.
• Built and optimized production ETL pipelines in IBM InfoSphere DataStage using SQL and PostgreSQL, processing millions of records daily to feed SAP BusinessObjects analytics dashboards, enabling self-service reporting access for 10+ enterprise departments.

• Advised executive leadership across 7 client engagements in banking, insurance, and e-payment sectors, delivering risk mitigation strategies that reduced audit findings by 30-40% on repeat engagements and helped clients pass subsequent audits.
• Performed SOX ITGC assessments across access management, change management, and data center/backup controls using the COBIT framework, delivering actionable remediation roadmaps for identified deficiencies.
• Identified and mitigated 20+ critical compliance risks through governance assessments and architecture reviews against ISO 27001 standards, safeguarding organizations from regulatory penalties, operational vulnerabilities, and reputational exposure.

Projects

• Built a budget-aware long-context pipeline (vLLM, Qwen3.6-35B) with a FAISS semantic answer cache and an LLM source verifier; on LongBench-v2 the cache cut total token usage by 49.3% and raised accuracy 3.4 points over the uncached baseline; retrieval beat truncation by 17 points on AA-LCR (54% vs 37%) and 13 on MRCR v2 (51% vs 38%).
• Engineered a benchmark-agnostic execution pipeline that counts prompt tokens exactly as vLLM does and routes each question to full context, middle truncation, or dense retrieval packing (Qwen3-Embedding-0.6B with FAISS, chunks ranked against the question and restored to source order) when the source exceeds the input budget.
• Designed lightweight task adapters so AA-LCR, MRCR v2, and LongBench-v2 share one pipeline with no benchmark-specific branching, and kept gold answers out of the cache and verifier so scoring stays leak-free.
• Showed that retrieval gains appear only once inputs exceed the model window: on MRCR v2 conversations of 533K and 1.06M tokens, accuracy rose from 27% to 60% and from 10% to 30%, and on the 107 LongBench-v2 questions that did not fit, from 41.1% to 56.1% (p = 0.0025).
• Automated experiments across a single 96 GB GPU server and a 4x L40S Slurm cluster with preflight routing reports, dry-run launchers, and 215 unit tests that run in about 3 seconds.

• Built a GenAI pipeline combining LLM tag generation (OpenAI API), Sentence-Transformers, KeyBERT, TextRank, NLTK, and NetworkX to extract, cluster, and link topics from unstructured documents. Achieved 0.97 weighted topic coherence on a 0-1 scale.
• Engineered an in-memory semantic knowledge graph (Python 3.13, sentence-transformers mpnet, NumPy) that turns local documents into topic-aware search, with LLM-backed tag extraction and statistical fallbacks (KeyBERT, TextRank).
• Designed a cosine-matching algorithm with configurable k and similarity threshold that links documents to topic centroids, then exported interactive PyVis HTML graphs and structured experiment logs for visualization and debugging.
• Evaluated extractors and thresholds: LLM tags at k=3, threshold 0.7 reached the best coherence (0.97), TextRank at k=5 or 7, threshold 0.3 reached full connectivity (largest-component ratio 1.00), and topic counts ranged from 32 to 475, exposing a clear granularity trade-off.

• Built a bioacoustics knowledge distillation pipeline with TensorFlow and Keras, converting audio signals into 2D Mel Spectrograms to train an EfficientNet Convolutional Neural Network (CNN) vision model on embeddings from Google's Perch v2 teacher model.
• Architected a multi-task EfficientNetB0 student with a shared backbone and two heads: a distillation head that mimics Perch v2 embeddings (1536 dimensions, mean squared error) and a classification head over 234 species trained on human-annotated soundscapes (binary cross-entropy), to identify overlapping bird calls in noisy Pantanal field recordings.
• Reached a leaderboard score of about 0.60 (best 0.603) after only 3 training epochs under Google Colab compute limits, with dual supervision acting as a regularizer against the teacher's biases.
• Documented the full pipeline in a commented, public starter notebook: 5-second windows at 32 kHz, 128-band mel spectrograms, Perch v2 embedding extraction, training on mixed clean and field-recording data, and Kaggle inference.

Publications

• B. Perk, C. Bayraktaroğlu, E. D. Doğu, F. S. Ali, and Ö. Ökasap, 2020 IEEE Symposium on Computers and Communications (ISCC), Rennes, France, pp. 1–6. IEEE Xplore · DOI: 10.1109/ISCC50000.2020.9219696
• Proposed, designed, and analyzed Joulin, a peer-to-peer energy marketplace built on smart contracts where peers produce, buy, and sell energy without a trusted third party; built a proof-of-concept on the Ethereum blockchain.
• Demonstrated usability, flexibility, and resiliency with low Ethereum gas costs and quick response times, and tested the smart contracts with security tools to confirm they resist outside manipulation.