7 projects · 2025–2026
During my internship at Cevizsoft, I noticed users constantly getting lost across 6 complex ERP modules and opening support tickets just to locate common workflows. To remove that friction, I built an AI assistant that allows natural-language navigation. I created a lightweight intent classification engine using n-gram and keyword matching that routes queries to the exact workflow without external LLM dependencies.
During my internship at Kokteyl, I set out to unify live match, team, and league telemetry from three fragmented external APIs into a single high-performance hub. I built a background .NET 8 Worker Service to poll and ingest the streams, batch-upserting records into PostgreSQL via Dapper to prevent table contention. On the frontend, I paired React and Vite with .NET Minimal APIs to deliver fast filtering and rapid query responses.
When building the corporate web platform for Şahlandı Yapı from the ground up, I wanted more than just an elegant showcase—I engineered a solid technical foundation. After designing the UI/UX, I implemented full hreflang routing across 6 locales and structured Schema.org JSON-LD data. Pairing that with a performance-optimized React architecture ensured technical SEO compliance and improved organic search visibility.
Seeing support teams drown in incoming tickets and risk missing critical escalations, I built an automated NLP decision engine in Python. I vectorized feedback using TF-IDF and trained a Logistic Regression classifier layered with heuristic urgency scoring. The system instantly isolates high-priority tickets and drafts automated initial responses, cutting triage time from hours to seconds.
My fascination with cybersecurity and the recurring risk of authorization regressions slipping through code reviews drove me to build this tool. Using Python's AST module, I created a static scanner that inspects route definitions across the codebase to flag missing access controls. I wrapped it into a GitHub Actions CI/CD workflow that automatically audits every pull request and comments with a risk score before code ever reaches production.
I built my graduation project on facial expression recognition and computer vision. I noticed models suffered severe accuracy drops under harsh lighting changes and skewed head poses in real-world conditions. To solve this, I redesigned the preprocessing and augmentation pipeline from scratch and benchmarked deep learning architectures, analyzing tradeoffs across metrics to achieve robust classification even under challenging conditions.
I didn't want a cookie-cutter template or a static bio page, so I coded this portfolio from scratch as a living software system that reflects my engineering mindset and design aesthetics. Built on TanStack Router and React, it incorporates client-side ASCII canvas rendering, low-level keyboard navigation, a hidden drop-down CLI terminal, and live telemetry signals, blending high performance with developer-centric craft.