Erik is available for hire
Hire ErikErik Harutyunyan
Verified Expert in Engineering
Artificial Intelligence (AI) Developer
Location
Yerevan, Armenia
Toptal Member Since
September 29, 2022
Erik is a master's student at the Technical University of Munich, studying mathematics in data science. During his last job at SuperAnnotate, he worked on various complex assignments, including edge detection, semantic segmentation, object detection, active learning, knowledge distillation, filter pruning, and deployment on edge device projects.
Portfolio
SuperAnnotate
Artificial Intelligence (AI), Computer Vision, Object Detection...
SuperAnnotate
Active Learning, Artificial Intelligence (AI), Python 3, Python, PyTorch...
SuperAnnotate
PyTorch, Python 3, TensorBoard, Deep Learning, Python, Data Science...
Experience
Availability
Part-time
Preferred Environment
Visual Studio Code (VS Code), Jupyter
The most amazing...
...thing I've developed is improving the speed of an edge detection model by around five times by continuous R&D in knowledge distillation and filter pruning.
Work Experience
Computer Vision Consultant
2021 - 2022
SuperAnnotate
- Advised and shared expertise with LiDAR point cloud data, guiding the development of the LiDAR annotation tool.
- Guided the video action recognition smart feature in the video editor tool.
- Led the real-time object tracking smart feature in the video editor tool.
- Proofread and commented on computer vision-related marketing articles and whitepapers.
Technologies: Artificial Intelligence (AI), Computer Vision, Object Detection, Semantic Segmentation, 3D Pose Estimation, LiDAR, Machine Learning, Deep Learning, Build Automation, Computer Vision Algorithms, Image Processing, Object Tracking, Jupyter, Machine Learning Operations (MLOps), Amazon SageMaker, Deep Neural Networks, NumPy, Neural Networks, Datasets, Generative Models
Computer Vision Researcher
2020 - 2021
SuperAnnotate
- Used state-of-the-art uncertainty estimation methods to develop the priority score feature for the tool that suggests to the annotator which images to annotate first.
- Improved the speed of an edge detection model by around five times applying state-of-the-art knowledge distillation and channel pruning methods, with only losing 1% in the F1 score.
- Researched and developed an improvement for the smart segmentation algorithm that lies in the core of the SuperAnnotate.
Technologies: Active Learning, Artificial Intelligence (AI), Python 3, Python, PyTorch, TensorBoard, Object Detection, Semantic Segmentation, 3D Pose Estimation, Visual Studio Code (VS Code), Machine Learning, Computer Vision, Deep Learning, OpenCV, Generative Adversarial Networks (GANs), Image Generation, Generalized Linear Model (GLM), Deployment, NVIDIA TensorRT, OpenVINO, Amazon Web Services (AWS), Google Cloud, Build Automation, Data Visualization, Computer Vision Algorithms, Image Processing, Object Tracking, Jupyter, Machine Learning Operations (MLOps), Amazon SageMaker, Deep Neural Networks, NumPy, Neural Networks, Reinforcement Learning, Datasets, Generative Models
Machine Learning Engineer
2018 - 2020
SuperAnnotate
- Employed OpenVINO and TensorRT to develop an integration covering from uploading raw data to the platform to training and subsequent deployment of object detection models to OAK-D and Jetson Nano devices, respectively.
- Developed analytics and visualization functions for the SDK to give insights into raw data and its annotations.
- Wrote medium articles and whitepapers about the cutting-edge projects done in SuperAnnotate.
Technologies: PyTorch, Python 3, TensorBoard, Deep Learning, Python, Data Science, Artificial Intelligence (AI), Dash, Visual Studio Code (VS Code), Dimensionality Reduction, Clustering, Machine Learning, Pandas, SciPy, Plotly, Amazon Web Services (AWS), Google Cloud, Build Automation, Data Visualization, Big Data Architecture, Image Processing, Jupyter, Machine Learning Operations (MLOps), Amazon SageMaker, Deep Neural Networks, NumPy, TensorFlow, Neural Networks, Datasets, PySpark
Data Scientist
2017 - 2018
Ucom
- Created an anomaly detection system using classical time series forecasting methods to predict the going down of internet spreading towers. The system helped to decrease the towers' downtime by four times on average.
- Developed tabular customer data of the company to predict the churn of each customer. The model reached around 85% accuracy when deployed in production.
- Used customer service usage data to provide a trustworthiness score that the company decided whether to sell mobile phones with monthly payments or demand the full price at once.
Technologies: Python 3, Scikit-learn, Pandas, SciPy, Plotly, Python, Data Science, Artificial Intelligence (AI), Dash, Probabilistic Graphical Models, Generalized Linear Model (GLM), Bayesian Statistics, Classification, Visual Studio Code (VS Code), Clustering, Machine Learning, PyTorch, Big Data, Data Visualization, Big Data Architecture, SQL, Jupyter, Machine Learning Operations (MLOps), NumPy, TensorFlow, Datasets, GPT, Natural Language Processing (NLP), Generative Pre-trained Transformers (GPT), PySpark
Experience
Active Learning for 3D Semantic Segmentation
https://github.com/ero1311/Pointnet2.ScanNetI took a famous 3D point cloud semantic segmentation model called Pointnet++. I developed an active learning algorithm using the well-known Monte Carlo dropout method to select new items from the unlabeled data pool. I organized and ran experiments on the 3D indoor scenes dataset, a ScanNetv2, in three granularity levels: scene level, segment level, and point level. In these three experiments, the algorithm selected whole scenes, segments from each scene, and points from each scene from the unlabeled pool. I could show that the algorithm significantly outperforms the random heuristic in all three experiments.
Consensus-based Optimization for Convolutional Neural Network
https://github.com/ero1311/cbo_implemenationsI implemented a novel class of gradient-free optimizers called: "Consensus-based Optimizers" in PyTorch. In addition, I wrote experiment codes that showed its successful applications on the famous Fashion-MNIST dataset to discover an autoencoder and a classifier with a fully convolutional neural network.
Analytics Functions for SuperAnnotate Python SDK
https://github.com/superannotateai/superannotate-python-sdkDeveloped several analytics functions for SuperAnnotate Python SDK that allowed the users to measure the quality of the annotations done on the platform.
FUNCTIONS INCLUDE:
• Consensus and benchmark that score annotators or groups of annotators on a small subset from the dataset to help the user select the ones whom they want to assign the labeling project.
• Embedding functions that produce 2D scatter plots for the dataset to give insights into the similarities and clusters present in the dataset.
• Annotation statistics and visualizations that include bar charts of classes, subclasses, attributes in the annotations, annotation times and counts for individual annotators, etc.
FUNCTIONS INCLUDE:
• Consensus and benchmark that score annotators or groups of annotators on a small subset from the dataset to help the user select the ones whom they want to assign the labeling project.
• Embedding functions that produce 2D scatter plots for the dataset to give insights into the similarities and clusters present in the dataset.
• Annotation statistics and visualizations that include bar charts of classes, subclasses, attributes in the annotations, annotation times and counts for individual annotators, etc.
From Raw Data to Model Deployment Pipelines
https://github.com/superannotateai/model-deployment-tutorialsDeveloped full pipelines in Google Colab that enable a user without coding skills to get a state-of-the-art object detection or semantic segmentation model deployed on an edge device from raw data. I developed these pipelines for NVIDIA Jetson and OAK-D devices. The pipelines include the following steps:
• Loading raw data to the SuperAnnotate platform
• Training a state-of-the-art model with annotated data
• Deploying the trained model to the edge device
• Loading raw data to the SuperAnnotate platform
• Training a state-of-the-art model with annotated data
• Deploying the trained model to the edge device
Education
2020 - 2022
Master's Degree in Mathematics in Data Science
The Technical University of Munich - Munich, Germany
2014 - 2018
Bachelor's Degree in Computer Science
The American University of Armenia - Yerevan, Armenia
Skills
Languages
Python 3, Python, SQL, R, C++
Libraries/APIs
PyTorch, OpenCV, NumPy, Scikit-learn, Pandas, SciPy, TensorFlow, PySpark
Tools
TensorBoard, Jupyter, Plotly, MATLAB, OpenVINO, Amazon SageMaker, ERDAS
Paradigms
Data Science
Platforms
Visual Studio Code (VS Code), Amazon Web Services (AWS)
Storage
Google Cloud
Other
Machine Learning, Computer Vision, Deep Learning, Optimization, Object Detection, Artificial Intelligence (AI), Generative Adversarial Networks (GANs), Image Generation, Active Learning, Probabilistic Graphical Models, Probability Theory, Statistics, Generalized Linear Model (GLM), Bayesian Statistics, Classification, Dimensionality Reduction, Clustering, Deployment, Semantic Segmentation, Build Automation, Big Data, Data Visualization, Big Data Architecture, Computer Vision Algorithms, Image Processing, Object Tracking, Machine Learning Operations (MLOps), Deep Neural Networks, Neural Networks, Datasets, Algorithms, Data Structures, OOP Designs, Mathematics, 3D, 3D Pose Estimation, Dash, NVIDIA TensorRT, GPT-2, Reinforcement Learning, Generative Models, Natural Language Processing (NLP), GPT, Generative Pre-trained Transformers (GPT), LiDAR
Collaboration That Works
How to Work with Toptal
Toptal matches you directly with global industry experts from our network in hours—not weeks or months.
1
Share your needs
Discuss your requirements and refine your scope in a call with a Toptal domain expert.
2
Choose your talent
Get a short list of expertly matched talent within 24 hours to review, interview, and choose from.
3
Start your risk-free talent trial
Work with your chosen talent on a trial basis for up to two weeks. Pay only if you decide to hire them.
Top talent is in high demand.
Start hiring