From staging layer to final output, I design systems that hold under real conditions. Then I add intelligence where it actually earns its place — not the other way around.
Each of these represents a full design loop — a real problem, a deliberate approach, a trade-off made consciously, and a result that is measured and defended.
Predictive modeling of sensor behavior in a live industrial production environment — Eliar Elektronik × Hacettepe AI Club
Predicting critical tank fill levels in a continuous paint production line — none of it clean, none of it static. Required handling real-time sensor signals, irregular valve states, and automated dosing commands generated by industrial machinery.
Refused naive cross-validation. Designed sequential validation to preserve temporal logic and prevent leakage. Built process-aware features around sequence dependencies — industrial signals carry memory. Iterated with LightGBM, focusing on stability under real production conditions, not just leaderboard gains.
Reliably outperformed baseline systems. Selected into the top 8 teams nationally. Invited to formally present and defend the architecture in Ankara before an academic and industrial professor jury.
AI-Powered Nutrition Intelligence Platform | Team Lead
People need more understandable, localized, and trustworthy nutrition guidance, especially when food information is fragmented across images, ingredient lists, and cultural food contexts.
Leading a multidisciplinary team in MVP definition, user persona development, feature prioritization, product discovery, and AI architecture planning. Contributing to the design of a multi-layer system combining computer vision, LLM-based reasoning, structured food knowledge, explainability, and safety-focused response generation.
Currently in early-stage development, with a validated MVP direction, responsible AI considerations, and a clearer product roadmap for Turkish and international users.
Developing an intelligent risk assessment architecture to optimize lending decisions.
Traditional credit scoring mechanisms often plateau at roughly 60% accuracy, relying heavily on simple linear thresholds. These baseline models fail to capture non-linear financial behaviors, resulting in excessive defaults and lost opportunities.
Engineered an ensemble machine learning pipeline incorporating XGBoost and Random Forest, orchestrated through rigorous cross-validation. Deployed the resulting serialized model as an interactive application via Hugging Face Spaces for immediate stakeholder access.
Elevated the baseline measurement from a stagnant 60% to a robust 75%, effectively proving the commercial value of ensemble trees in high-stakes financial risk modeling.
A retrieval-augmented pipeline executing precise database queries over complex schemas — Global AI Hub × Akbank
LLMs consistently hallucinate SQL structures when untethered. Without rigorous grounding in a defined schema context, generated analytics queries fail silently, projecting high confidence on corrupted data.
Built a FAISS vector retrieval layer mapping to specific normalized entity embeddings. Developed structured, strict prompt templating boundaries to inject context safely, and cut latency via prefix query filtering.
Engineered and deployed a fully functional MVP within one week, dropping systemic hallucination rates drastically while reducing retrieval latency by 35%.
End-to-end analytical architecture mapping raw staging areas through 3NF up to a modeled data mart — EPAM Systems
Analytical stakeholders were executing demanding queries directly against fragile operational databases. The ecosystem lacked separation of concerns, mutation tracking, or recovery strategies for failed pipelines.
Migrated infrastructure to a strict Staging → 3NF → Data Mart paradigm. Enforced Slowly Changing Dimensions (SCD Type 1/2) and built heavily idempotent ETL jobs backed by robust transaction controls.
Successfully decoupled BI layers from operational schema fluctuations. Eliminated silent reporting failures by establishing fully auditable, rerunnable ingestion pipelines.
Long-range climate visualization and correlation analytics engine.
The bottleneck wasn't the data — it was access. Translating decades of raw atmospheric and temperature datasets into immediate, actionable insights for non-technical researchers without losing statistical rigor.
Built highly reusable Python data pipelines leveraging Pandas for ingestion and Scikit-learn for correlation modeling. Architected a live frontend employing Dash and Plotly for deep interactive visualization.
Delivered a thesis-style analytics application capable of rendering long-range correlations dynamically via parameter injection from the UI.
Data engineering should be invisible when functioning properly. The structural decisions made in the foundational ingestion layer irrevocably dictate the integrity of every subsequent analytic.
Every ETL workflow makes implicit promises downstream. I make those promises explicit by integrating idempotency, transaction commits, and structured error boundaries at deployment, not post-failure.
Raw ingestion, normalized 3NF, and dimensional models aren't bureaucratic filler—they represent strict ownership boundaries. Collapsing layers merely to accelerate immediate delivery is borrowing heavily against future stability.
Any intelligent model is fundamentally capped by the reliability of its input features. I architect AI capabilities exclusively atop sanitized, well-modeled schemas. Reliable engineering renders intelligence trustworthy.
Sophisticated analytical work holds zero leverage if misunderstood by decision-makers. I prioritize translating pipeline latency, predictive confidence, and data quality into universally resonated business narratives.
Languages and frameworks rotate over time. The structural methodologies remain constant. Below are my core areas of execution.
Where research meets production constraints.
Paper: A Portable Provenance Schema for Auditable Clinical NLP
Presented an applied AI research paper focused on provenance, auditability, and reliability in clinical NLP systems at ICADA 2026.
Recognized as an AI Content Creator and awarded "Rising Voice of the Month". Passionate about translating technically dense concepts—like ETL mechanics and deployment nuances—for learners and non-experts.
Graduate of the SAP Young Professionals Program. Holding verified credentials including SAP Certified Associate – Implementation Consultant and SAP Certified – SAP Generative AI Developer. Profound exposure to end-to-end business models via SAP S/4HANA.
Certifications, test scores, and artifacts that back the work.
My foundational reflex is identical regardless of the domain: I must interpret the macro-system before mutating the micro-parts. This instinct guides my data architecture, but it also dictates my work outside the terminal.
I operate extensively in narrative storytelling and voice delivery. While it appears to be separate from engineering, it relies on the same core principle: absorbing complexity and delivering it so the intent is understood without friction.
Explaining systems is not an accessory — it is a constraint. If something cannot be made clear without distortion, the design itself is incomplete.
I am drawn to work at the intersection of architecture and execution — where understanding the full system matters as much as delivering a clean result.
Whether I’m designing a data pipeline or shaping a narrative, the objective is the same: preserve intent, reduce noise, and make the outcome usable.
Currently accessible for opportunities traversing data engineering, analytics architecture, and applied intelligence layers. Remote or Istanbul based.