Applied AI, data systems, and automation.

Designed a controlled synthetic-fact framework to measure English-to-Turkish factual transfer separately from reaffirmation and relearning. Built bilingual pre/post probing, adaptation branches, and controls for frequency, entity distribution, contamination, and prompt paraphrases.

Designed a compact Space XES extension for structured coordinates, CRS, floor/level, and discrete location semantics. Implemented PM4Py preprocessing for spatial-column detection, validation, type normalization, and CRS handling; tested it across three datasets covering 736 events and seven routes.

Built a Dockerized PostgreSQL environment and modular Python QueryBuilder for clinical parameter extraction. Implemented Bronze-Silver-Gold ETL layers, unit standardization, plausibility checks, indexes, and a Gold-layer SOFA-score pipeline.

Led end-to-end operations for a private 10-person on-chain market-intelligence project while members retained independent execution and risk decisions. Built Python and Rust automation for blockchain data collection, token monitoring, analytics, and profitability estimation; the group collectively realized approximately $2.5M in net P&L between 2023 and 2025.

Compared five traditional annotation-error methods with 15 LLM probability approaches across five models on VariErr NLI. The strongest DeepSeek-R1 few-shot setup reached 0.2756 Average Precision, about 20% above the best traditional baseline.

Built a reusable, responsive developer-portfolio template with YAML-driven content, validated data, theme support, subtle motion, and separate landing and portfolio routes.

Built a controlled research pipeline comparing four prompting strategies for reducing gender bias in LLM-generated educational text. Across 300 experimental trials, Few-Shot + Verification and System Prompt achieved 84.1% bias reduction while preserving semantic quality.