Linkedin

Github

Email

"Hi! I am Alper Erten, a machine learning engineer with experience across machine learning, deep learning, and reinforcement learning, including applications in natural language processing, computer vision, robotics, and energy systems optimization."

"Hi! I am Alper Erten, a machine learning engineer with experience across machine learning, deep learning, and reinforcement learning, including applications in natural language processing, computer vision, robotics, and energy systems optimization."

My projects span from foundational ML algorithms to cutting-edge deep RL and real-time scene understanding systems.

In my prior life, I led a team of engineers helping companies reduce their carbon footprint through energy modeling and analysis tool development.

My projects span from foundational ML algorithms to cutting-edge deep RL and real-time scene understanding systems.

In my prior life, I led a team of engineers helping companies reduce their carbon footprint through energy modeling and analysis tool development.

Reinforcement Learning

Reinforcement Learning

Deep Reinforcement Learning in Continuous Action Spaces: Lunar Lander

This study benchmarks two continuous-action deep reinforcement learning algorithms, Deep Deterministic Policy Gradient (DDPG) and Twin Delayed Deep Deterministic Policy Gradient (TD3), in the Lunar Lander v3 environment. The evaluation examines how stability mechanisms, exploration strategies, and update schedules influence policy learning. Results show that while DDPG can solve the task, it suffers from high reward variance and slower convergence, whereas TD3 consistently achieves higher rewards with faster, more stable learning, establishing TD3 as a strong baseline for continuous control tasks.

PYTORCH

GYMNASIUM

PYGAME

Deep Reinforcement Learning in Continuous Action Spaces: Lunar Lander

This study benchmarks two continuous-action deep reinforcement learning algorithms, Deep Deterministic Policy Gradient (DDPG) and Twin Delayed Deep Deterministic Policy Gradient (TD3), in the Lunar Lander v3 environment. The evaluation examines how stability mechanisms, exploration strategies, and update schedules influence policy learning. Results show that while DDPG can solve the task, it suffers from high reward variance and slower convergence, whereas TD3 consistently achieves higher rewards with faster, more stable learning, establishing TD3 as a strong baseline for continuous control tasks.

PYTORCH

GYMNASIUM

PYGAME

Deep Reinforcement Learning in Continuous Action Spaces: Lunar Lander

This study benchmarks two continuous-action deep reinforcement learning algorithms, Deep Deterministic Policy Gradient (DDPG) and Twin Delayed Deep Deterministic Policy Gradient (TD3), in the Lunar Lander v3 environment. The evaluation examines how stability mechanisms, exploration strategies, and update schedules influence policy learning. Results show that while DDPG can solve the task, it suffers from high reward variance and slower convergence, whereas TD3 consistently achieves higher rewards with faster, more stable learning, establishing TD3 as a strong baseline for continuous control tasks.

PYTORCH

GYMNASIUM

PYGAME

Coordinated Task Execution in Multi-Agent Reinforcement Learning: Overcooked-AI

This study benchmarks three value-based MARL algorithms including Deep Q-Networks (DQN), Value Decomposition Networks (VDN), and QMIX, on cooperative tasks in the Overcooked-AI environment. Using two layouts with varying coordination demands, it evaluates the impact of value factorization, reward shaping, and training stability under centralized training and decentralized execution (CTDE). Results show QMIX consistently outperforms DQN and VDN, especially in high-coordination scenarios like the Coordination Ring, achieving near optimal soup deliveries with minimal errors. Findings highlight QMIX’s superior ability to model cooperative credit assignment, positioning it as a strong baseline for structured coordination tasks.

PYTORCH

OPENAI GYM

PYGAME

Coordinated Task Execution in Multi-Agent Reinforcement Learning: Overcooked-AI

This study benchmarks three value-based MARL algorithms including Deep Q-Networks (DQN), Value Decomposition Networks (VDN), and QMIX, on cooperative tasks in the Overcooked-AI environment. Using two layouts with varying coordination demands, it evaluates the impact of value factorization, reward shaping, and training stability under centralized training and decentralized execution (CTDE). Results show QMIX consistently outperforms DQN and VDN, especially in high-coordination scenarios like the Coordination Ring, achieving near optimal soup deliveries with minimal errors. Findings highlight QMIX’s superior ability to model cooperative credit assignment, positioning it as a strong baseline for structured coordination tasks.

PYTORCH

OPENAI GYM

PYGAME

Coordinated Task Execution in Multi-Agent Reinforcement Learning: Overcooked-AI

This study benchmarks three value-based MARL algorithms including Deep Q-Networks (DQN), Value Decomposition Networks (VDN), and QMIX, on cooperative tasks in the Overcooked-AI environment. Using two layouts with varying coordination demands, it evaluates the impact of value factorization, reward shaping, and training stability under centralized training and decentralized execution (CTDE). Results show QMIX consistently outperforms DQN and VDN, especially in high-coordination scenarios like the Coordination Ring, achieving near optimal soup deliveries with minimal errors. Findings highlight QMIX’s superior ability to model cooperative credit assignment, positioning it as a strong baseline for structured coordination tasks.

PYTORCH

OPENAI GYM

PYGAME

Deep Learning

Deep Learning

Scene Understanding for Robotics

This project developed a real-time scene graph generation pipeline in ROS 2 using YOLOv11 and synchronized RGB-D data. It detects objects, infers spatial relationships, and integrates LiDAR-based depth sensing to build semantic maps. Enhancements include reference surface detection, relationship filtering, and stability optimizations. The system supports assistive robotics applications by enabling structured, symbolic scene understanding from live visual input.

ROS2

OPENCV

YOLO11

OPEN3D

Scene Understanding for Robotics

This project developed a real-time scene graph generation pipeline in ROS 2 using YOLOv11 and synchronized RGB-D data. It detects objects, infers spatial relationships, and integrates LiDAR-based depth sensing to build semantic maps. Enhancements include reference surface detection, relationship filtering, and stability optimizations. The system supports assistive robotics applications by enabling structured, symbolic scene understanding from live visual input.

ROS2

OPENCV

YOLO11

OPEN3D

Scene Understanding for Robotics

This project developed a real-time scene graph generation pipeline in ROS 2 using YOLOv11 and synchronized RGB-D data. It detects objects, infers spatial relationships, and integrates LiDAR-based depth sensing to build semantic maps. Enhancements include reference surface detection, relationship filtering, and stability optimizations. The system supports assistive robotics applications by enabling structured, symbolic scene understanding from live visual input.

ROS2

OPENCV

YOLO11

OPEN3D

Semi-Supervised Learning with Lightweight Image Classifiers

This study explores the application of the FixMatch algorithm to lightweight image classifiers, aiming to optimize semi-supervised learning for resource-constrained environments. Leveraging the CIFAR-10 dataset, the study evaluates the performance of three architectures: ResNet-8, ResNet-20, and a custom OptVGGNet model. The findings highlight the potential of lightweight models in achieving efficient learning while reducing computational demands and data-labeling costs in semi-supervised setups.

PYTORCH

TORCHVISION

SCIKIT-LEARN

Semi-Supervised Learning with Lightweight Image Classifiers

This study explores the application of the FixMatch algorithm to lightweight image classifiers, aiming to optimize semi-supervised learning for resource-constrained environments. Leveraging the CIFAR-10 dataset, the study evaluates the performance of three architectures: ResNet-8, ResNet-20, and a custom OptVGGNet model. The findings highlight the potential of lightweight models in achieving efficient learning while reducing computational demands and data-labeling costs in semi-supervised setups.

PYTORCH

TORCHVISION

SCIKIT-LEARN

Semi-Supervised Learning with Lightweight Image Classifiers

This study explores the application of the FixMatch algorithm to lightweight image classifiers, aiming to optimize semi-supervised learning for resource-constrained environments. Leveraging the CIFAR-10 dataset, the study evaluates the performance of three architectures: ResNet-8, ResNet-20, and a custom OptVGGNet model. The findings highlight the potential of lightweight models in achieving efficient learning while reducing computational demands and data-labeling costs in semi-supervised setups.

PYTORCH

TORCHVISION

SCIKIT-LEARN

Deep Learning Model Interpretability & Style Transfer

This project explores deep learning interpretability and style transfer using lightweight CNNs. Techniques like saliency maps, Grad-CAM, and class visualizations reveal how models make decisions. Neural style transfer blends image content with artistic styles using feature-based losses. Built with PyTorch and Captum, the work emphasizes gradient-based visualization and creativity through efficient, interpretable neural network applications.

PYTORCH

CAPTUM

NUMPY

Deep Learning Model Interpretability & Style Transfer

This project explores deep learning interpretability and style transfer using lightweight CNNs. Techniques like saliency maps, Grad-CAM, and class visualizations reveal how models make decisions. Neural style transfer blends image content with artistic styles using feature-based losses. Built with PyTorch and Captum, the work emphasizes gradient-based visualization and creativity through efficient, interpretable neural network applications.

PYTORCH

CAPTUM

NUMPY

Deep Learning Model Interpretability & Style Transfer

This project explores deep learning interpretability and style transfer using lightweight CNNs. Techniques like saliency maps, Grad-CAM, and class visualizations reveal how models make decisions. Neural style transfer blends image content with artistic styles using feature-based losses. Built with PyTorch and Captum, the work emphasizes gradient-based visualization and creativity through efficient, interpretable neural network applications.

PYTORCH

CAPTUM

NUMPY

NLP Machine Translation

This project explores German-to-English neural machine translation using Sequence-to-Sequence (Seq2Seq) models and Transformers. It begins with RNN and LSTM-based encoder-decoder architectures enhanced with attention, followed by implementing Transformer models with multi-head self-attention and masked decoding. Through extensive evaluation and tuning, Transformer models achieved superior fluency and performance, highlighting the shift from traditional recurrent networks to modern attention-based architectures in NLP.

PYTORCH

SCIKIT-LEARN

NLP Machine Translation

This project explores German-to-English neural machine translation using Sequence-to-Sequence (Seq2Seq) models and Transformers. It begins with RNN and LSTM-based encoder-decoder architectures enhanced with attention, followed by implementing Transformer models with multi-head self-attention and masked decoding. Through extensive evaluation and tuning, Transformer models achieved superior fluency and performance, highlighting the shift from traditional recurrent networks to modern attention-based architectures in NLP.

PYTORCH

SCIKIT-LEARN

NLP Machine Translation

This project explores German-to-English neural machine translation using Sequence-to-Sequence (Seq2Seq) models and Transformers. It begins with RNN and LSTM-based encoder-decoder architectures enhanced with attention, followed by implementing Transformer models with multi-head self-attention and masked decoding. Through extensive evaluation and tuning, Transformer models achieved superior fluency and performance, highlighting the shift from traditional recurrent networks to modern attention-based architectures in NLP.

PYTORCH

SCIKIT-LEARN

Machine Learning

Machine Learning

Benchmarking Supervised Learning Models on Real World Datasets

This study evaluates Support Vector Machines, k-Nearest Neighbors, and Neural Networks on two contrasting datasets—phishing detection and wine quality prediction. It benchmarks model performance across accuracy, speed, and tuning sensitivity. Findings highlight that SVMs offer the most consistent results, KNN provides simplicity, and neural networks require careful tuning but can achieve competitive performance with sufficient data.

PYTORCH

SCIKIT-LEARN

Benchmarking Supervised Learning Models on Real World Datasets

This study evaluates Support Vector Machines, k-Nearest Neighbors, and Neural Networks on two contrasting datasets—phishing detection and wine quality prediction. It benchmarks model performance across accuracy, speed, and tuning sensitivity. Findings highlight that SVMs offer the most consistent results, KNN provides simplicity, and neural networks require careful tuning but can achieve competitive performance with sufficient data.

PYTORCH

SCIKIT-LEARN

Benchmarking Supervised Learning Models on Real World Datasets

This study evaluates Support Vector Machines, k-Nearest Neighbors, and Neural Networks on two contrasting datasets—phishing detection and wine quality prediction. It benchmarks model performance across accuracy, speed, and tuning sensitivity. Findings highlight that SVMs offer the most consistent results, KNN provides simplicity, and neural networks require careful tuning but can achieve competitive performance with sufficient data.

PYTORCH

SCIKIT-LEARN

Dimensionality Reduction and Clustering for Real-World Classification Tasks

This project explores unsupervised learning techniques—PCA, ICA, Random Projections, K-Means, and EM—to reduce feature space and uncover data structure in phishing and wine datasets. Results show ICA offered the best compression-performance tradeoff, while clustering aided exploratory insights. Neural networks trained on ICA-transformed data preserved accuracy while reducing training time, demonstrating practical benefits of dimensionality reduction.

PYTORCH

SCIKIT-LEARN

Dimensionality Reduction and Clustering for Real-World Classification Tasks

This project explores unsupervised learning techniques—PCA, ICA, Random Projections, K-Means, and EM—to reduce feature space and uncover data structure in phishing and wine datasets. Results show ICA offered the best compression-performance tradeoff, while clustering aided exploratory insights. Neural networks trained on ICA-transformed data preserved accuracy while reducing training time, demonstrating practical benefits of dimensionality reduction.

PYTORCH

SCIKIT-LEARN

Dimensionality Reduction and Clustering for Real-World Classification Tasks

This project explores unsupervised learning techniques—PCA, ICA, Random Projections, K-Means, and EM—to reduce feature space and uncover data structure in phishing and wine datasets. Results show ICA offered the best compression-performance tradeoff, while clustering aided exploratory insights. Neural networks trained on ICA-transformed data preserved accuracy while reducing training time, demonstrating practical benefits of dimensionality reduction.

PYTORCH

SCIKIT-LEARN

App Development

App Development

Offerwise: Android App Development

Offerwise is a modular Android app for evaluating job offers using weighted comparisons of salary, benefits, and cost-of-living factors. Built with MVVM architecture, the app emphasizes clean UI separation and persistent storage. It ranks offers using a customizable scoring algorithm. The project highlights full-stack mobile engineering, data modeling, and test-driven development for real-world decision support.

ANDROID STUDIO

JAVA

ANDROID JETPACK

Offerwise: Android App Development

Offerwise is a modular Android app for evaluating job offers using weighted comparisons of salary, benefits, and cost-of-living factors. Built with MVVM architecture, the app emphasizes clean UI separation and persistent storage. It ranks offers using a customizable scoring algorithm. The project highlights full-stack mobile engineering, data modeling, and test-driven development for real-world decision support.

ANDROID STUDIO

JAVA

ANDROID JETPACK

Offerwise: Android App Development

Offerwise is a modular Android app for evaluating job offers using weighted comparisons of salary, benefits, and cost-of-living factors. Built with MVVM architecture, the app emphasizes clean UI separation and persistent storage. It ranks offers using a customizable scoring algorithm. The project highlights full-stack mobile engineering, data modeling, and test-driven development for real-world decision support.

ANDROID STUDIO

JAVA

ANDROID JETPACK

StoreSmart: Carbon-Aware Energy Storage Control Optimizer

StoreSmart is a Python-based control tool for optimizing battery storage to reduce carbon emissions. It integrates real-time carbon intensity data, PV forecasts, and grid rates to schedule environmentally favorable charge/discharge events. With a modular architecture and rule-based logic, it simulates annual system performance, aiding energy managers and researchers in designing carbon-aware storage strategies.

PYTHON

PANDAS

CARBON OPTIMIZATION

StoreSmart: Carbon-Aware Energy Storage Control Optimizer

StoreSmart is a Python-based control tool for optimizing battery storage to reduce carbon emissions. It integrates real-time carbon intensity data, PV forecasts, and grid rates to schedule environmentally favorable charge/discharge events. With a modular architecture and rule-based logic, it simulates annual system performance, aiding energy managers and researchers in designing carbon-aware storage strategies.

PYTHON

PANDAS

CARBON OPTIMIZATION

StoreSmart: Carbon-Aware Energy Storage Control Optimizer

StoreSmart is a Python-based control tool for optimizing battery storage to reduce carbon emissions. It integrates real-time carbon intensity data, PV forecasts, and grid rates to schedule environmentally favorable charge/discharge events. With a modular architecture and rule-based logic, it simulates annual system performance, aiding energy managers and researchers in designing carbon-aware storage strategies.

PYTHON

PANDAS

CARBON OPTIMIZATION

DecarbPath Modeler: Portfolio Decarbonization Planning Tool

The DecarbPath Modeler is a scalable simulation framework for evaluating greenhouse gas reduction strategies across building portfolios. It uses modular scenario logic to model electrification, efficiency, and renewables over time. Aligned with DOE guidelines, the tool enables emissions forecasting, target benchmarking, and capital planning support through structured inputs, temporal modeling, and scenario-based decision visualization. The tool received a software copyright license from the U.S. Department of Energy Patent Counsel.

DATA ANALYSIS

SCENARIO MODELING

DecarbPath Modeler: Portfolio Decarbonization Planning Tool

The DecarbPath Modeler is a scalable simulation framework for evaluating greenhouse gas reduction strategies across building portfolios. It uses modular scenario logic to model electrification, efficiency, and renewables over time. Aligned with DOE guidelines, the tool enables emissions forecasting, target benchmarking, and capital planning support through structured inputs, temporal modeling, and scenario-based decision visualization. The tool received a software copyright license from the U.S. Department of Energy Patent Counsel.

DATA ANALYSIS

SCENARIO MODELING

DecarbPath Modeler: Portfolio Decarbonization Planning Tool

The DecarbPath Modeler is a scalable simulation framework for evaluating greenhouse gas reduction strategies across building portfolios. It uses modular scenario logic to model electrification, efficiency, and renewables over time. Aligned with DOE guidelines, the tool enables emissions forecasting, target benchmarking, and capital planning support through structured inputs, temporal modeling, and scenario-based decision visualization. The tool received a software copyright license from the U.S. Department of Energy Patent Counsel.

DATA ANALYSIS

SCENARIO MODELING

Software Stack

Software Stack

Python

Java

Android Studio

Pytorch

Library

ROS 2

OpenCV

Library

Scikit-learn

Library

Pandas

Library

Numpy

Library

© Alper Erten. 2025

CREATED WITH LOVE IN SAN FRANCISCO