Afshin Shahrestani

Projects & Publications

Technical projects and research contributions.

Personal projects, professional work, and research from over the years.

View projects Publications

Projects

Personal And Professional Projects

RAG Document Retrieval System

A Retrieval-Augmented Generation (RAG) pipeline for querying medical patient records using vector similarity search and language models.

  • Vector-based document retrieval using FAISS and HuggingFace embeddings.
  • LLM-powered question answering with Microsoft Phi-3-mini model.
  • Synthetic patient data generation using Faker and LLM.
  • 11 demonstration scenarios covering single-patient queries, multi-chunk retrieval, structured output formatting, and more.

Stack: Python, LangChain, FAISS, HuggingFace, PyTorch, Phi-3-mini.

Source Code →

Pavement Crack Image Generator

Developed Deep Convolutional Generative Adversarial Networks (DCGAN), Wasserstein GAN (WGAN), WGAN with Gradient Penalty (WGAN_GP), and Wasserstein Conditional GAN (cGAN) with Gradient Penalty (C_WGAN_GP) to generate synthetic images of pavement cracks.

  • Built GAN architectures from scratch for synthetic crack image generation.
  • Used to augment existing datasets and improve ML model robustness.
  • Facilitates research in pavement maintenance and repair.

Stack: Python, PyTorch, DCGAN, WGAN, Conditional GAN.

Source Code →

Road Image Object Annotation and Detection

Deep Learning object detection models, including YOLOv5 and Faster R-CNN, were used to detect different objects such as traffic signs, traffic roads, pavement lane marking, and pavement cracks in road images.

  • Gathered images from Canadian roads and public datasets.
  • Annotated using CVAT to be compatible with YOLO and COCO formats.
  • Trained and evaluated multiple detection architectures.

Stack: Python, YOLOv5, Faster R-CNN, CVAT, OpenCV.

Source Code →

Better Exam - Accessible Exam Platform

An exam hosting service for the visually impaired, built using Azure Cognitive Speech Services. Ranked top 50 in Azure AI Hackathon 2021.

  • Text-to-Speech reads questions aloud; Speech-to-Text captures student answers.
  • Intuitive touch gestures and voice navigation for accessibility.
  • Teachers create exams via web tool; students access via QR codes or links.
  • Answers compiled into PDF and emailed to teachers automatically.

Stack: Python, Flask, Azure Cognitive Services, JavaScript, HTML/CSS.

Source Code →

Power Consumption Trend Prediction & Fraud Detection

Used deep learning models such as Autoencoders, RNNs, MLPs, and data analysis techniques like missing value interpolation, clustering, and decision trees to develop Trend Prediction and Fraud Detection software.

  • Preprocessing including data normalization, standard scaling, and outlier removal.
  • Achieved 92%+ accuracy in both trend prediction and fraud detection.
  • Incorporated into 2 software products sold to Mashhad Electrical Utility Company.

Stack: Python, TensorFlow, Autoencoders, RNN, Scikit-learn.

Source Code →

Publications

Peer-reviewed Research

45 citations · h-index: 2 · Google Scholar →

IEEE Transactions on Intelligent Transportation Systems · 2025

Conditional Pavement Crack Data Generation for Selective Data Augmentation Using GAN

Developed a Conditional GAN (cGAN) model enabling selective data augmentation for pavement crack classification. The approach allows targeted generation of specific crack types to address class imbalance issues in road surface datasets.

  • Implemented conditional generation for targeted crack type synthesis.
  • Addressed class imbalance in road defect datasets.
  • Improved classification accuracy through selective augmentation.
IEEE T-ITS →

TAC Conference · 2024 · 2 citations

Pavement Crack Image Generation and Data Balancing Using Deep Adversarial Generative Networks

Presented DCGAN and WGAN models for synthetic road image generation, demonstrating how deep adversarial networks can effectively balance imbalanced pavement crack datasets and enhance classification model performance.

  • Developed DCGAN and WGAN architectures for road imagery.
  • Demonstrated effective data balancing strategies.
  • Enhanced downstream classification performance.
TAC Conference →

University of British Columbia · 2024 · 1 citation

Using Wasserstein Distance and Conditionality for Stable and Selective Data Augmentation in Pavement Crack Analysis

Master's thesis exploring the use of Wasserstein distance (WGAN) and conditional generation techniques for stable training and targeted data augmentation in pavement crack analysis, achieving 15% improvement in crack classification accuracy.

  • Explored Wasserstein distance for stable GAN training.
  • Implemented conditional generation for selective augmentation.
  • Achieved 15% improvement in crack classification accuracy.
UBC Library →

Electric Power Systems Research · 2022 · 36 citations

Online Electricity Theft Detection Framework for Large-Scale Smart Grid Data

Developed an AI-based framework for real-time fraud detection in smart grid power consumption data. The model achieved 90% accuracy in detecting electricity theft patterns across large-scale datasets.

  • Built real-time fraud detection for smart grid systems.
  • Achieved 90% accuracy on large-scale datasets.
  • Most cited publication with 36 citations.
EPSR Journal →

IEEE IoT Conference · 2021 · 6 citations

Filter-Based Time-Series Anomaly Detection in AMI Using AI Approaches

Proposed AI-based anomaly detection methods for Advanced Metering Infrastructure (AMI) using time-series filtering techniques. The approach achieved 92% accuracy in trend prediction for smart meter data.

  • Developed filter-based anomaly detection for AMI systems.
  • Achieved 92% accuracy in trend prediction.
  • Applied to real-world smart meter datasets.
IEEE Xplore →