Dilan Üzümcü

Data Scientist with 3+ years of experience in applied machine learning, grounded in statistics (BSc Statistics, Hacettepe University). Builds and evaluates models on tabular, geospatial, image and video data using gradient boosting, vision-language models and LLM-based pipelines, with an emphasis on sound validation and measurable business impact (+15% targeting effectiveness, ~10% higher segmentation accuracy, ~2× pipeline throughput). Has worked on projects deployed on AWS and Azure using Docker, MLflow and CI/CD.

Author of an IEEE UBMK 2025 paper and data scientist on a TÜBİTAK-supported R&D project. Open to remote, hybrid or on-site roles.

Data Scientist (Contract) at BilgeAdam Teknoloji (2026-05 – 2026-08)

• Multimodal Video Understanding & Automated Topic-Labeling System: Built a multimodal video-understanding pipeline (keyframes, speech recognition, OCR, audio/object cues) that generates topic labels for retrieval and classification without manual annotation.
• Designed a cost-aware cascade: a lightweight classifier handles ~90% of videos and low-confidence cases escalate to retrieval + VLM/LLM reasoning, cutting expensive inference by 20% and roughly doubling end-to-end throughput.
• Packaged the pipeline with Docker and CI/CD for deployment on Azure, added a daily retraining loop and RAG components.
• Tools: Python, Hugging Face Transformers, Whisper, OCR, FAISS, Docker, Azure.

Data Scientist (Full-Time) at Mapin Data (2023-07 – 2026-02)

• Replaced manual expert evaluations with a scoring and segmentation system for 8,678 HoReCa venues, converting menus, reviews, social media and mobile-location signals into quantitative scores; improved location-based product targeting effectiveness by 15%.
• Engineered 200+ features and 16 composite scores, trained a weakly supervised multi-task LightGBM/CatBoost model (4-class segmentation, binary tiering, capture-probability regression), improving production segmentation accuracy by ~10% over the baseline.
• Rebuilt venue-photo scoring after VLM embeddings underperformed: LLM-generated captions → text embeddings → visual score, raising F1 from 0.90 to 0.95.
• Built a PostgreSQL/PostGIS geospatial data lakehouse (500+ variables, 5+ sources) and automated model training and tracking with SageMaker, S3 and MLflow: 2× faster queries and 12% less manual analysis time.
• Worked with product teams in a B2B SaaS setting to define model requirements, including the latency vs. accuracy trade-off.
• Tools: Python, SQL, PostgreSQL/PostGIS, LightGBM, CatBoost, scikit-learn, AWS (SageMaker, S3), MLflow.

Data Science Mentor & Statistical Consultant (Freelance, Part-Time) at Saksı Kampüs (2024-01 – 2025-11)

• Mentored 20+ junior data scientists from problem framing to EDA, modeling and evaluation (Python, R), with a 90% project-completion rate; worked alongside a full-time role at Mapin Data.

Statistician / Data Analyst (Freelance) at Ankara, Türkiye (2022-06 – 2022-08)

• Delivered statistical modeling, experimental design, hypothesis testing, regression analysis and technical reporting for research and business clients using R and Python.

Ankara Chapter Lead (Volunteer) at Omdena (2023-06 – Present)

• Lead local-chapter AI projects with 100+ participants, mentor junior data scientists and run case-based learning sessions.

BSc in Statistics – Hacettepe University (2018-09 – 2022-05)