Istanbul, Türkiye
I work across the stack. RTL design and computer architecture through hardware-aware software and production systems. At Amazon, I architected an agentic Text-to-SQL platform that became the first Amazon Turkey product to expand internationally. At ITU's Synergy Computer Architecture Research Group, I designed TinyNPU (a systolic array NPU) in SystemVerilog with a mixed-precision code generation pipeline from PyTorch FX graphs to hardware runtime. I'm interested in the intersection of hardware architecture and efficient inference: how systems are designed from the ground up to run models well.
Architected and led end-to-end development of an agentic AI Text-to-SQL platform that became the first Amazon Turkey product to expand internationally. Owned and led the development of: Redshift data warehouse design, backend services, UI. Collaborated on the agentic query generation layer and the ETL pipeline integrating Amazon's core data systems into Incepta's Redshift cluster.
Built a churn prediction model that directly reduced churn by 200bps, along with a fully automated retraining and deployment pipeline for continuous model maintenance. Developed an LLM-based customer sentiment tracking system to surface real-time signals from user feedback at scale. Designed optimized Redshift caching layers to replace long-running queries, eliminating bottlenecks in business-critical analysis workflows. Built & maintained analytical dashboards and AWS data pipelines that expanded the team's self-serve data capabilities. (first and third most used dashboards in the organization)
Contributed to PhD research on LLM post-training quantization, evaluating quantization strategies across transformer components using SmoothQuant, TensorRT-LLM, and vLLM. Designed and implemented TinyNPU, a stolic array NPU in SystemVerilog with mixed-precision support (W4A4 / W8A8 / W16A16), targeting efficient on-device deep learning inference. Built a compiler pipeline from PyTorch FX graphs to NPU runtime inputs. Build RISC-V implementation on FPGA that passes the official test set.
Developed a GPT-3 powered microservice chatbot for website navigation using LangChain, ChromaDB, and Django. Built the frontend for the funds crowdfunding website using Next.js and contributed Go/JS features across the platform, including i18n localization for English expansion.
Built a customer feedback classifier achieving 96%+ accuracy across 20+ categories using FastText and Zemberek, directly integrated into customer experience workflows. Developed an NLP pipeline using topic modeling (Gensim) and the Apriori algorithm to identify products driving negative feedback and delivery issues. Built and maintained a Selenium-based web crawler for competitive price monitoring, and collaborated with finance, security, and CX teams on ad-hoc data analysis. Created and maintained Tableau dashboards and evaluated Superset as a cost-reduction alternative to existing BI tooling.