Browse My Recent

Projects

Monocular Depth Estimation & Obstacle Detection

An end-to-end single-camera obstacle-detection pipeline that estimates depth from a single image and flags nearby obstacles. Key innovations include:

Combined Depth Anything V2 and MiDaS with relative-to-metric depth alignment and distance-thresholded obstacle masking; benchmarked 4 models on the NYU Depth V2 dataset.
Selected Depth Anything V2 ViT-S as the best accuracy-speed trade-off, running at 42 FPS on an A100 GPU.
Fine-tuned Depth Anything V2 ViT-S on 30K RGB-D pairs via hyperparameter search and freezing-strategy ablations, reducing AbsRel by 40% and improving obstacle-detection F1 from 0.48 to 0.66.

Speech-Based Health Assessment with Multimodal Deep Learning

A multimodal multi-task deep learning system that assesses emotional and health-related states from speech by combining acoustic and linguistic features. Key innovations include:

Combining Whisper-base (acoustic) and BERT (linguistic) representations via cross-attention for joint audio-text modeling.
Multi-task learning setup to jointly predict multiple emotional/health-related states from a single model.
Achieved a 0.6517 mean macro F1 on the held-out test set, improving on the baseline by 0.0445.

Multilingual Idiom Identification (Turkish & Italian)

A project focused on token-level identification of idiomatic multi-word expressions (MWEs) in Turkish and Italian sentences using a transformer-based sequence labeling pipeline. Key innovations include:

Training on a combined multilingual corpus to improve coverage and generalization across languages.
Language-specific BERT encoders with feature fusion from the last hidden layers, followed by a BiLSTM to capture long-range context.
CRF decoding to enforce coherent BIO tag sequences and recover complete idiom spans.
Two-phase optimization: task-specific training with frozen BERT, then gradual layer-wise unfreezing with reduced learning rates to fine-tune stably.
Achieved 90 overall F1 in the course competition, demonstrating strong performance on morphologically rich language data.

Distributed Image Similarity Search

A big data project focused on detecting duplicate and near-duplicate images at scale for automated plagiarism/copyright checks, using deep image embeddings and distributed processing. Key innovations include:

Extracting semantic image representations with DINOv3 and comparing them using cosine similarity for robust near-duplicate detection.
Precomputing and L2-normalizing embeddings, then storing them in Parquet for fast, scalable Spark workflows.
Designing a distributed pipeline in Apache Spark and Google Cloud Dataproc that supports top-k retrieval and threshold-based flagging.
Benchmarking across multiple dataset sizes (up to ~157K images) and showing that distributed similarity search stays fast as the archive grows, enabling near real-time retrieval.

Predictive Analytics for E-commerce Product Replenishment

A project aimed at forecasting customer repurchase timing in e-commerce using machine learning techniques. Key innovations include:

Utilization of XGBoost and CatBoost models for predicting repurchase windows.
Advanced feature engineering, including temporal analysis and product grouping via Jaccard similarity.
Application of over 1 million transactional records to optimize inventory management and supply chain operations.

Emotion Representation in TTS Models

A project focused on improving the emotional expressiveness of Text-to-Speech (TTS) systems by integrating emotion embeddings into the synthesis pipeline. Key innovations include:

Leveraging StyleTTS2's diffusion-based architecture to generate high-quality, emotionally stable speech.
Incorporating an emotion discriminator trained on datasets like RAVDESS to refine emotional accuracy in synthesized audio.
Application of diverse datasets (RAVDESS, LibriTTS) for robust emotional modeling.
Introducing emotion labels into the TTS pipeline and optimizing their integration using cross-entropy loss to enhance emotional fidelity in synthesized speech.

Database Design and Implementation

A project aimed at designing a robust database system for music streaming platforms to manage vast amounts of data related to artists, albums, songs, playlists, and user interactions. Key features include:

Development of a RESTful API using Flask-RESTX with SQLite backend, JWT-based authentication, and modular architecture for secure and efficient data management.
Comprehensive CRUD operations for managing entities like users, playlists, songs, and albums, ensuring seamless interaction with the database.
Integration of advanced relational modeling to maintain data integrity and enable complex queries like tracking user engagement or calculating total listening time per album.

Berke Kurt

About Me

Education

Work Experience

AI Intern — Netmera

Research Intern — Özyeğin University

AI Intern — Turkcell

Areas of Interest

Natural Language Processing

Computer Vision

Speech and Audio Processing

Information Retrieval & RAG

Deep Learning

Big Data

MLOps

Data Mining

Database Systems

Algorithms

Cognitive Neuroscience

Game Theory