cv

Ph.D., Computer Science

Florida International University

• Reviewer: NeurIPS, ICML, ICLR, ITW, ISIT, T-IFS
• Thesis: A Stochastic Approach to Privacy-Preserving Distributed Parameter Estimation

M.Sc., Control Engineering

Sharif University of Technology

• Thesis: A Fuzzy Probabilistic Belief Framework for Uncertainty Propagation in Dynamical System Reliability

B.Sc., Control Engineering

Tabriz University

• Thesis: Quantitative Application of Non-Cooperative Game Theory Models for Stock Market Prediction

Graduate Research Assistant

Florida International University

• Maximally Negatively Correlated Quantizers in Distributed Learning (CorBin-FL)
  • Introduced correlated distributed quantization for federated learning, improving accuracy over prior methods.
  • Validated on CIFAR10, FEMNIST, Shakespeare, and Sent140 (CNN, ResNet18, LSTM, Transformer) on multiple A100s, achieving high communication efficiency by transmitting one bit per model weight.
• Explanation-Preserving Augmentation for Semi-Supervised Graph Representation Learning
  • Developed a semantics-preserving graph augmentation technique leveraging GNN explainers, with theoretical analysis of its gains over prior methods.
  • Validated on MUTAG, Benzene, and PROTEINS, achieving up to 7.83% improvement in classification accuracy.
• Discrete Denoising Diffusion Model (DiGress) and Application in Graph Explainability
  • Applied Discrete Denoising Diffusion (DiGress) to ProxyExplainer to improve GNN explainer accuracy.
  • Conditioned the diffusion model on explainer subgraphs to enable in-distribution graph generation, preserving explainable structure while reconstructing the rest.
• Denoising Diffusion Implicit Model (DDIM) for Time Series Prediction
  • Developed a conditional DDIM with an attention-based UNet for time-series analysis and generation.
  • Encoded volume and price movements as multi-channel RGB conditioning to capture temporal dependencies.
• Agentic-SysID: Autonomous Multi-Agent System Identification
  • Built a multi-agent pipeline (Intake, Modeler, Estimator, Validation) converting a natural-language plant description into a control-ready model with uncertainty bounds and a validity region, under a constrained budget.
  • Designed a physics-first fidelity ladder escalating from white-box to grey-box or black-box surrogate only on validation evidence, routed through a guarded, budget-enforced Plant API; demonstrated on a pendulum-with-friction plant that breaks the white-box model.
• Multi-Agent Sports Game Analysis (LangGraph)
  • Built a LangGraph-orchestrated multi-agent pipeline for NBA/MLB game analysis with parallel adversarial debates (win-probability, total-points), each converged by a judge and cross-validated by a head analyst.
  • Closed the loop with a post-game accuracy review that writes semantically-retrieved lessons to a ChromaDB memory store per agent role, across a mixed OpenAI/Anthropic LLM routing layer with SQLite persistence.

Research and Development Manager

Tunnel Sadd Ariana

• AI-driven stock market prediction: web data extraction, feature selection, preprocessing (FFT, LOWESS), and LSTM-based stock-market prediction.

Senior Electrical Engineer

Tunnel Sadd Ariana

• Managed the implementation of industrial automation: PLC programming and SCADA design for irrigation-system water-flow control and monitoring.
• Supervised electrical and industrial-automation section of two comprehensive World Bank projects.

Research Fellow

K.N. Toosi University

• Researched gradient-based optimization methods: applied a second-order gradient-based technique to solve optimal control problems in hybrid dynamical systems.
• Developed advanced control strategies for bipedal and monopod robots: extended the hybrid minimum principle (HMP) to control a monopod robot for maximum jumping height.

eBay ML Challenge: NLP-Driven Auto Part Classification

• NLP for information extraction: combined Named Entity Recognition and inference-based methods to pair vehicle components from free-text descriptions.
• Data preprocessing: cleaned and filtered text, applied spaCy-based importance summarization, and built augmented, auto-labeled training sets.
• Two-stage BERT fine-tuning: one model extracts initial labels directly; a second infers missing fields from context, improving accuracy and recall.

Detection of DDoS Attacks using SVM and Logistic Regression

• Built an SVM/Logistic-Regression DDoS classifier on the CIC-DDoS2019 dataset (SMOTE-balanced) and on self-generated traffic captured via hping3/Wireshark.
• Achieved high accuracy and low false-positive rates across both datasets, showing real-world applicability.

A Genetic Algorithm and Tabu Search Method for Supply Chain Problems

• Developed a Genetic Algorithm and Tabu Search for capacity-constrained location-routing, overcoming exact-method computational limits for large-scale Mixed Integer Programming (MIP) in supply-chain problems; presented at the 2nd International Congress on Engineering Technology and Innovation.