Projects
AI and robotics first, but also the ML, full-stack and mobile work that got me here.
2026Independent research
Measuring the demonstration-count vs. success-rate curve for SmolVLA, LoRA fine-tuned on unseen LIBERO tasks, on a single 8 GB consumer GPU. Nearly all of the value arrives in the first five demonstrations, and so does catastrophic forgetting.
66.7% · success @ 5 demos
2026TEKNOFEST 2026 · Team bugbuster · Finalist
TEKNOFEST 2026 Artificial Intelligence in Aviation. I owned the GPS-free positioning task: RAFT optical flow + homography with a keyframe ladder estimates the aircraft’s displacement from its downward camera. Full rehearsal: 2256/2256 frames, 3.54 m mean error.
3.54 m · mean error, full run
2026TEKNOFEST 2026 · 5G & AI Smart Road Safety · Team BiDatalar
Identifies a vehicle (body type, plate, colour) and detects driver-caused violations from a single night-time pass through the windshield. Built on one rule: nothing that cannot be proven is reported. Result: zero false positives.
1.00 · precision
2026Independent project · public dataset
Neural dead reckoning for a fixed-wing UAV: an LSTM predicts one-second displacements from 19 GPS-free sensor channels. After 4.5 minutes without GPS it is still within ~85 m, about 38× better than classical dead reckoning.
44.4 m · mean error, 4.5 min
2026TEKNOFEST 2026 · Unmanned Ground Vehicle
ROS 2 Humble autonomy stack for a 4×4 unmanned ground vehicle on Jetson Orin NX: SLAM, EKF sensor fusion, Nav2, a slalom planner and a finite-state-machine mission manager, developed in Gazebo simulation first.
ROS 2 · Humble
2026Open-source tutorial
A hands-on state-estimation library: KF and EKF implemented from scratch in NumPy, with seven progressive examples from 1D tracking to sensor fusion, EKF radar and a real drone flight-log case study.
7 · worked examples
2026TEKNOFEST 2026 · Healthcare AI
Missense genetic variants → pathogenic / benign, with ~55% missing features and a train/test prior shift. Leakage-controlled validation, Optuna-tuned gradient-boosting ensemble and SHAP explanations.
SHAP · explainability
2026TUSAŞ Lift Up · Graduation project
Graduation project in the TUSAŞ Lift Up program. Five regressors compared under three cross-validation schemes on a Taguchi L9 design; the RSM model reaches CV R² = 0.978. Shipped as a desktop app, standalone .exe and Flask REST API.
R² 0.978 · cross-validated
2026
Classifying lung cancer risk from patient symptoms with CatBoost, validated with 5-fold cross-validation.
AUC 0.94 · 5-fold CV
2026Hackathon project
Two-sided skill-exchange marketplace with trust, credit-balance and fairness scores: Spring Boot backend, Flutter app and an ML credit-scoring regression.
2025
Eight mini-projects across computer vision and NLP: pneumonia detection with transfer learning, YOLOv8 vehicle tracking, CNN classification, LSTM and a RAG pipeline with FAISS and Gemini.
2025Finalist
Customer-churn prediction model built for the ING Hub Datathon, where the team reached the finals.
2025
Neighbourhood super-app: events, groups, real-time chat over STOMP, marketplace, local pros with reviews and maps. Flutter client with a Spring Boot backend.
2025
REST APIs on a layered controller–service–repository architecture, Keycloak security and Swagger / OpenAPI docs.
2025TEKNOFEST Air Defense
Desktop control interface for an air-defense system concept: live camera feed, mission-stage controls and telemetry.
2025
Sixteen classic algorithms (convolution, Canny, histogram equalisation, morphology) written in raw NumPy, no OpenCV filter calls, behind a PyQt5 interface.
16 · algorithms
2024
A collection of small cross-platform apps built while learning Dart and Flutter.
2024
25 SwiftUI components, animations and mini-apps consolidated into one organised showcase.
2023
Language fundamentals (events, delegates, generics) and Windows Forms desktop applications.