Esra Cesur

Developed and deployed a comprehensive web-based AI diagnostic platform that automates disease classification and anomaly detection across chest, bone, dental, and ophthalmology images. Implemented automated PDF reporting and integrated Gemini for clinical decision support and second-opinion guidance.

Designed and trained staged pipelines combining EfficientNet baselines with a two-phase Vision Transformer approach (binary No-DR vs DR, then severity grading). Optimized rare-class sensitivity with Focal Loss and achieved a Kaggle test accuracy of 0.86075 with K-Fold ensembling.

Designed and implemented an end-to-end AI pipeline combining speech recognition (OpenAI Whisper), text-based emotion classification done with custom built RoBERTa models, and text-to-speech (Edge-TTS) to generate synchronized talking avatars. Developed custom viseme-based facial animation, integrated multilingual emotion models (Turkish/English), and deployed a Flask-based web application.

Developed a Retrieval-Augmented Generation (RAG) application that transforms any YouTube video into an interactive study companion. Integrated Whisper for transcription, SentenceTransformer + ChromaDB for semantic retrieval, and Groq (LLaMA 3.3) for contextual Q&A, quiz, and flashcard generation within a Flask–TailwindCSS architecture backed by SQLite.

Engineered two secure keyless entry prototypes: (1) 433 MHz RF + RFID proximity with AES, (2) BLE with AES + hopping code and custom mobile app. Implemented proximity ignition, app control, and 3D-printed enclosures for functional demos.

Built an image stylization pipeline in MATLAB that segments objects with K-Means, refines masks via morphology, and renders lightweight depth-like effects for an artistic 3D look.

Designed an interactive Power BI report with KPIs (revenue, orders, AOV), customer segmentation, category trends, and regional analysis using star-schema modeling and custom DAX measures.

Developed a hybrid approach combining a custom CNN and transfer learning (VGG16, MobileNet, ResNet50) to classify Round, Scoop, and V-Neck images. Addressed class imbalance with SMOTE and achieved the most balanced results with ResNet50.

Built a CNN with three convolutional blocks, batch normalization, and dropout to recognize 24 ASL letters (static). Achieved ~99.9% test accuracy with smooth learning curves and strong generalization.

Implemented a classification pipeline using Spotify audio features to predict track mood (Happy, Sad, Angry, Chill). Benchmarked Random Forest, Naïve Bayes, and XGBoost; selected XGBoost for the best balance of accuracy and class consistency.

Developed a Streamlit application to estimate academic performance from lifestyle and prior records using Gradient Boosting Regressor with GridSearchCV and 5-fold CV, achieving R² ≈ 0.86 and offering interactive feature-importance views.

Fine-tuned T5-small (with BART as an alternative) on CNN/DailyMail to generate concise news summaries. Evaluated outputs with ROUGE-1/2/L and packaged scripts for reproducible training and inference.

Conducted an end-to-end EDA to identify key drivers of satisfaction—delays, service quality, and demographics—using robust preprocessing, feature engineering, and interactive visual analytics.

Performed statistical and visual analysis on engine telemetry with outlier handling, RPM banding, temperature deltas, and interactive 3D/correlation heatmaps to surface condition insights.

Developed a Power BI dashboard featuring KPIs, geospatial mapping, mass classification, and decade/hemisphere analyses using custom DAX and interactive slicers.