Projects
Competitions
Lumina
First place, Mentorship phase
Open-source study platform that uses retrieval-augmented generation to turn uploaded PDFs into quizzes, flashcards and study guides.
What I built
- Designed a 40+ table PostgreSQL schema (SQLAlchemy, Alembic) with pgvector and ChromaDB for vector search, with course-level isolation enforced in the method signature so no query can reach another course's material.
- Provisioned AWS ECS Fargate, RDS, S3 with CloudFront, an Application Load Balancer and Secrets Manager in Terraform; built continuous integration and deployment in GitHub Actions with OpenID Connect login to AWS, Pytest, Ruff, Playwright end-to-end tests and automated database backup and restore.
- Conceived and built Reverse Quiz: students answer open-ended questions and the AI checks their answers against the source material, turning passive recall into active retrieval practice.
AWS, Terraform, Docker, GitHub Actions, PostgreSQL, pgvector, ChromaDB, SQLAlchemy, Alembic, FastAPI, Python
WellWell
Medication-safety app that scans medicine labels, flags interactions and duplicated active ingredients, and sends offline dose reminders.
What I built
- Reads brand, generic name, active ingredient, dose and expiry from a label photo and checks them against a drug database; adds reminders with streak tracking, caregiver sharing and an emergency QR card.
- Flutter app (Riverpod, go_router) and a .NET 8 vertical-slice API with PostgreSQL, Redis and JSON Web Token authentication with refresh-token rotation and reuse detection. Every safety decision is deterministic; GPT-4o-mini only writes the explanations.
- 168 automated tests (107 unit, 61 integration); the backend is deployed live through continuous integration.
Flutter, Dart, .NET 8, PostgreSQL, Redis, GPT-4o-mini
TOYGAR
Turkic-themed 3D action-defense game: defend a nomadic otağ (a ruler's tent) during a traditional wedding.
What I built
- Culturally inspired gameplay loop with localized resources and custom defensive mechanics.
- Dynamic day-night audio.
Unreal Engine 5, Blueprints
Projects
Neural networks from scratch
An autograd engine and character-level language models built from first principles, from a single neuron to a WaveNet-style model.
What I built
- Autograd engine with backpropagation for every operation, verified three ways (as a single tanh operation, as decomposed primitives, and numerically), plus a pure-Python renderer that draws computation graphs as scalable vector graphics (SVG).
- Character-level language models trained on a Turkish names dataset: a bigram model, then a multilayer perceptron after Bengio et al. (2003), then a hierarchical WaveNet-style model.
- Deriving backpropagation by hand through batch normalization and cross-entropy, without autograd.
Python, PyTorch
Bilkent Campus Simulation
Interactive 2D campus-management game in Java and JavaFX, built on a model-view-controller architecture.
What I built
Java, JavaFX
Other
The ceramics industry in Türkiye's industrialisation: Toprak Holding
Group research paper on the Turkish economy after 1980, traced through one company: the rise of Toprak Holding and Toprak Seramik in Bozüyük, and the collapse that followed its move into banking.
What I built
- Built from newspaper archives, audiovisual sources, and an interview with a company executive from the period.
- Published in Bilkent University's institutional repository under CC BY-NC-ND 4.0. Written in Turkish.