Vcamp

• Served as Team Captain for the TEKNOFEST Havacılıkta Yapay Zeka (AI in Aviation) Competition, defining the system architecture and coordinating the object-detection, position-estimation, and reference-object-matching workstreams.
• Designed a hybrid visual odometry pipeline combining ORB/optical flow and RAFT-based motion estimation with confidence-aware arbitration and iterative EKF fusion, achieving approximately 7.5 FPS on custom drone footage.
• Built an RGB/thermal aerial reference-object matching pipeline and fine-tuned YOLOv8 over seven training rounds to reach 0.749 mAP50 at approximately 36 ms per frame inference.
• Led the four-class UAV object-detection subsystem, including dataset architecture, training strategy, and validation criteria, on a dataset containing more than 35K images.
• Designed the competition's unified real-time client, integrating the three subsystems behind a modular inference pipeline with a competition-compliant JSON submission format.

Ravelos

• Built an intermittent-demand forecasting system for a 20K+ SKU spare-parts catalogue with approximately 1.4M transaction lines and 94% zero cells at daily grain.
• Improved forecast accuracy from 14% to approximately 55% across 26 benchmarked variants, reducing WAPE to 0.543 from a 0.628 last-year baseline.
• Designed a two-stage LightGBM hurdle architecture for demand occurrence and non-zero quantity, with SBC-adaptive thresholds and leakage-safe recursive multi-step inference.
• Engineered approximately 25 features, including Croston, ADI/CV², rolling and lag windows, and zero-run streaks; used Numba JIT, cutoff-masked feature generation, and fold-local SBC labels.
• Built evaluation and monitoring infrastructure with monthly vintage fact tables, FastAPI, Grafana dashboards, restatement detection, and champion/challenger shadow evaluation.
• Optimized the multi-model pipeline with joblib process parallelism, float32 downcasting, vectorized rank/correlation computation, and O(1) lookup maps replacing O(P) scans, with parity tests preserving output.

TUARIQ

• Audited production readiness across backups, migrations, deployment, health checks, testing, dependency consistency, and the repository's handling of secrets and large artifacts.
• Developing a phased, test-gated improvement plan covering reliable backups, migration recovery, deployment rollback, meaningful readiness checks, CI testing, and observability.
• Analyzed synchronous API workflows that block a single-worker API for 4–65 seconds per file; scoped worker/queue-based execution and measurement as performance priorities.
• Documented repository and service refactoring priorities, keeping behavior-changing work separate from hygiene and structural improvements.

Sompo Sigorta

• Built an end-to-end cross-sell propensity pipeline in KNIME with period-aware data-quality screening, rolling quantile binning against the previous month's breakpoints, and target-preserving category consolidation.
• Benchmarked decision tree, GBM, and WoE-based logistic regression under forward/backward feature selection and Gini elimination, evaluated across separate time windows.
• Developed a loss-adjuster amount regression under strict leakage control using time-ordered splits and train-only imputation; handled a heavily right-skewed target via log transformation, reducing skewness from 5.45 to 0.26, and evaluated Tweedie, Gamma, Poisson, and two-stage hurdle architectures.
• Componentized the pipeline into a single parameter-driven run generalizable to new datasets and authored internal reports on workflow automation tooling, including n8n and KNIME Business Hub, covering KVKK, access control, and deployment.