Ardavan Sassani

Summary

Skills

• Computer Vision: Object Detection, Image Segmentation, Facial Recognition
• Machine Learning: Supervised Learning, Unsupervised Learning, Reinforcement Learning, Data Warehousing, Predictive Modeling, Feature Engineering, Time Series Analysis, Big Data Analytics, Cluster Analysis
• Deep Learning: Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Generative Adversarial Networks (GANs), LLMs & RAG.
• Programming Languages: Python, MATLAB, C#
• Tools and Libraries: TensorFlow, Keras, PyTorch, PySpark, DataBricks, OpenCV, Scikit-Learn, NumPy, Pandas, Matplotlib, Seaborn, MLOps, WinFlow, Google VertexAI Studio.
• Environmental Engineering: Water Quality, GPS-X, WaterCAD, AutoCAD, ArcGIS, QGIS, BioWin, WaterGEMS, Civil 3D, EPA Net.

Education

• Doctor of Philosophy in Computer Science, Georgia State University, 2025
• Master of Science in Computer Science, Georgia State University, 2023
• Master of Engineering in Environmental Engineering, Azad University of Tehran, 2010
• Bachelor in Environmental Engineering, Azad University of Tehran, 2007

Experience

• Research Assistant in ML and Data Eng. , Georgia State University, 2020-Present

  • Computer Vision: Proficient in small object detection within satellite imagery using cutting-edge frameworks including U-NET, Fast RCNN, SAM, and YOLO. Demonstrated proficiency in addressing complex challenges in remote sensing and image analysis. Applied spatial attention mechanisms for enhanced accuracy in object localization. Strong programming skills in languages such as Python and proficiency in utilizing relevant libraries and tools (e.g., OpenCV, TensorFlow, PyTorch).
  • Data Engineering: Implemented an auto data collection/integration system to provide an updated data warehouse using C\# and PostgreSQL on the Azure cloud platform. It collects data from multiple sources on the web (web scraping).
  • Intelligent Early Warning System: Designed and built an intelligent early warning system to predict arboviral diseases (Dengue, Yellow fever, and Zika) outbreaks using machine learning algorithms based on land and weather conditions in tropical regions.
  • Online Framework for Calculating Shortest Path: Designed and implemented an online framework for calculating the shortest path for a vehicle routing system based on multiple distance formats interfaced by a rich REST API service written with ASP.Net.
  • Collaborated with cross-functional teams to analyze and interpret data, and provide actionable insights.
  • Mentored undergraduate students in machine learning and computer vision projects.

• Data Scientist, 3M inc., 2022-2023

  • Machine Learning End-to-End Pipeline: Designed and implemented a fully automated machine learning end-to-end pipeline to detect phantom inventory events based on historical sales data for each item/retail store using Azure Databricks platform with microservices design and improved the model performance by 25\% in terms of F1\_Score with XGBoost and LightGBM models.
  • Computer Vision: Implemented computer vision algorithms (U-NET, Fast RCNN, SAM, YOLO) to detect and count objects on shelves, mitigating phantom inventory events. Utilized Python, OpenCV, TensorFlow, and PyTorch for efficient and accurate small object detection in real time.
  • Azure Databricks and Spark Framework: Utilized Azure Databricks tool-sets and the Spark framework to improve productivity and performance.
  • Time Series Analysis: Developed a time series model to forecast the sales of each item/retail store for the next 30 days using Azure Databricks platform with PySpark and utilizing the Prophet model to improve the accuracy.
  • Feature Engineering: Designed and implemented a feature engineering pipeline to extract features from the sales data for each item/retail store.
  • Big Data Analytics, Data Integration and Transformation: Collaborated with the data engineers to create fully automated pipelines to integrate, update, and transform data across several retailers using Azure Data Lake Gen2 and Function App services to minimize manual tasks.
  • Software Engineering: Designed and implemented a reusable predictive time-series package to forecast product sales based on historical data using time-series models (Decomposition, Prophet, etc.) and clustering algorithms to improve the accuracy of prediction.
  • System Design:Collaborated with cross-functional teams to analyze and interpret data, and provide actionable insights.

• Full-stack Developer, TechClass Oy, 2017-2019

  • Designed and Implemented REST-full service using Entity Framework and .NetCore libraries coded with C\# and MSSQL database engine.
  • Optimized the authentication and authorization flow using the OAuth technique provided by Google, LinkedIn, GitHub, and Facebook.
  • Implemented facial recognition system for identity verification in a mobile application.
  • Contributed to the front-end as a developer to implement a fast and user-friendly web application using the ReactJS framework.

• Full-stack Developer, SFI Group, 2015-2017

  • Designed and implemented a multi-platform (web, desktop) integrated customized production management using ASP.Net and Angular.
  • Collaborated with the back-end team as a developer and system designer in an agile environment.
  • Collaborated with architectures and system designers to improve the user experience based on use cases and business logic.
  • Data was stored in MSSQL, and clients had access to it through a REST API service and an Angular web application. Quality control officers and their managers use this software to keep track of all produced parts, create customized reports, and estimate packing and shipping schedules.

• Process Engineer, IRGES consulting company, 2013-2016

  • Intelligent Bioreactor Controller: Designed and implemented a bioreactor controller using a PLC and an ANN to to optimize the population of microorganisms that consume alcoholic substances or other organic materials in a bioreactor. The system was designed to be user-friendly and to provide real-time monitoring and control of the bioreactor.
    • Process Model (MATLAB) to simulate the growth of microorganisms in a bioreactor.
    • ANN for Intelligent Control (MATLAB) to optimize the growth of microorganisms.
    • Database and Queries (MS SQL) to store and retrieve data.
    • PLC and Control Logic (C#) to read sensors and control actuators.
    • Windows Form Application (C#) as a user interface to monitor and control the bioreactor.
    • Electrical Interfaces and Controllers to connect the PLC to the bioreactor sensors and actuators.
  • Performed as responsible engineer preparing process and I\&C drawings including PFDs, P\&IDs, piping, and layouts.
  • Prepared projects proposals.

• Environmental Engineer, Zolal_Ab Sanat, 2009-2013

  • Designed advanced treatment systems including Reverse Osmosis (RO), Membrane Bio Reactor (MBR), and Ultra Filtration (UF) to recycle wastewater effluent.
  • Designed water and wastewater facilities including pumping, storing, and transmission systems. This includes pipe sizing, selecting pumps, valves, tanks size, and equipment.
  • Prepared drawings, and specifications to support project design packages including but not limited to, plant layout, yard piping, piping, and process layouts.

Projects

Artificial Intelligence and Machine Learning

Intelligent Early Warning System

I have developed an intelligent early warning system (IEWS) that utilizes machine learning and computer vision algorithms to forecast outbreaks of arboviral diseases such as Dengue, Yellow Fever, and Zika. The system analyzes environmental and meteorological conditions in tropical regions, and it also incorporates image processing to detect and quantify specific mosquito species in photographs submitted by the local community. I have used Python, OpenCV, TensorFlow, and PyTorch for efficient and accurate small object detection in real time. Read more about this project.

Online Framework for Calculating Shortest Path

I have designed and implemented an online framework for calculating the shortest path for a vehicle routing system based on multiple distance formats interfaced by a rich REST API service written with ASP.Net. I have utilized deep reinforcement learning algorithms to simulate and optimize the shortest path for a vehicle routing system. The system is capable of handling multiple distance formats such as Euclidean, Manhattan, and real road distance. The system is also capable of handling vehicle capacity and time window constraints.

Small Object Detection In Satellite Imageries

Semantic segmentation for images can be defined as partitioning and classifying the image into meaningful parts and classifying each part at the pixel level into one of the pre-defined classes.

One issue of this project is the small training dataset size with the total number of training images of 25 object-labeled images. To overcome this problem, we use U-Net architecture, an extended form of Fully Connected Network FCN. The main idea of this approach is to use a CNN as a powerful feature extractor while replacing the fully connected layers with convolution ones to output spatial maps instead of classification scores. Those maps are up-sampled to produce dense per-pixel output. This method allows training CNN in the end-to-end manner for segmentation with input images of arbitrary sizes. Read more

Short Time Earthquake Prediction

Accurately predicting earthquakes is critical for several reasons:

• Early Warning: Providing early warning to the public can save lives and reduce property damage.
• Emergency Response: Emergency response teams can prepare for the aftermath of an earthquake.
• Infrastructure Protection: Engineers can design buildings and infrastructure to withstand seismic activity.

Predicting earthquakes, especially in the short term (hours or days in advance), is notoriously difficult. This stems from the inherent complexity of earthquakes themselves.

Our Project's Approach

This project tackles this challenge by simplifying the problem. We will use a set of techniques to convert the prediction of these continuous variables into a categorical classification problem. In simpler terms, instead of pinpointing the exact magnitude, location, and time, we will aim to categorize them into predefined ranges (e.g., high magnitude, low magnitude). This approach can potentially improve the accuracy of short-term earthquake prediction. Read more

Web Development

Pakoob web application(under construction)

Pakoob is a web application designed to streamline your outdoor adventures. It combines carpooling services with mapping tools, making it easier than ever to:
• Find a ride to your next outdoor adventure.
• Connect with fellow outdoor enthusiasts.
• Plan your route using built-in mapping tools.