Developed a real-time facial spoofing detection system by comparing a classical machine learning pipeline (combining feature extractors like LBP with PCA and an SVM) against a MobileNetV2-based deep learning model. Trained on the combined LCC FASD and CASIA datasets, the deep learning approach achieved 97.4% accuracy and a 91.7% F1-score. The MobileNetV2 model also demonstrated superior robustness and generalization, outperforming the classical pipeline by over 20 percentage points in accuracy when evaluated on the unseen iBeta video dataset, effectively highlighting the advantages of modern deep learning techniques.

This project involved a detailed evaluation of large language models using a custom-built Retrieval-Augmented Generation (RAG) framework. I developed an end-to-end pipeline, vectorizing over 235,000 Q&A pairs from the AI Medical Chatbot dataset with the all-MiniLM-L6-v2 model for context retrieval. An automated evaluation framework, using gte-large-en-v1.5 to generate semantic similarity scores, was created to quantitatively assess query relevance and document faithfulness. This system was used to benchmark Mistral-7B, Llama-3-8B, and Qwen2-7B, identifying Mistral-7B as the top-performing model with a 10% higher composite score and superior inference speed.

Developed a multi-class emotion classification model for Persian text by benchmarking several fine-tuned RoBERTa variants on the ArmanEmo dataset. To handle the dataset’s over 7,000 noisy social media texts, I engineered a robust preprocessing pipeline featuring text normalization, diacritic removal, and artifact handling. The final model, XLM-RoBERTa-Large, was selected for its superior performance, achieving a 75.3% Macro F1-Score—a greater than 13x improvement over the baseline—and ranking as a top project in its course. The project concluded with a detailed error analysis that identified key failure modes, such as sentence ambiguity and word-level biases.

Developed a Persian extractive question-answering system by benchmarking and fine-tuning ParsBERT and XLM-RoBERTa-Large models on the PersianQA dataset. A robust text normalization pipeline using Parsivar was implemented to handle language-specific complexities. To provide a comprehensive assessment, a dual evaluation framework was created, combining standard metrics (Exact Match and F1-score) with a custom semantic similarity score based on all-MiniLM-L6-v2 embeddings. After fine-tuning, the XLM-RoBERTa-Large model was identified as the top performer, achieving an 85.15% F1-score and a 70.96% Exact Match score, outperforming the fine-tuned ParsBERT (72.67% F1, 57.84% EM). Both models surpassed the original XLM-RoBERTa (84.81% F1, 70.40% EM) and ParsBERT (70.06% F1, 53.55% EM) baselines reported by the dataset’s authors.

Developed a Persian Automatic Speech Recognition (ASR) system by benchmarking three models—Whisper-Large-V3, Seamless-M4T-V2-Large, and stt_fa_fastconformer_hybrid_large—on a Persian speech dataset sourced from YouTube. The FastConformer model was subsequently fine-tuned using the NVIDIA NeMo toolkit. This process substantially improved its performance, reducing the Word Error Rate (WER) from 0.74 to 0.37 and the Character Error Rate (CER) from 0.59 to 0.19. The resulting model outperformed the best baseline, Whisper-Large-V3 (WER of 0.56), demonstrating the effectiveness of language-specific adaptation for speech recognition tasks.

Developed a comprehensive system to assist university students with course selection, architected on a scalable backend using Django and PostgreSQL and deployed in a production-ready Docker environment with automated CI/CD. A key component was a multithreaded web crawler that automated the ingestion of over 1,000 classes per cycle, achieving an >11x speedup and bypassing logins with a KNN-based CAPTCHA solver that reached 96% accuracy. The system featured social functionalities like real-time chat, a BERT-based semantic search engine for content discovery, and instant course availability notifications via Email, SMS, and Telegram, all accessible through a RESTful API.

Developed a reinforcement learning agent to solve the classic Mountain Car problem by implementing and comparing on-policy (SARSA) and off-policy (Q-Learning) algorithms with linear function approximation. A key innovation in this project was engineering a compact feature set and applying L2 regularization to stabilize learning. This optimization proved highly effective, dramatically accelerating convergence for the SARSA agent from over 1000 episodes down to just 2 episodes.

Designed and implemented a VHDL-based hardware accelerator for CNNs optimizing performance for real-time applications. Conducted extensive testing to ensure accuracy and reliability. This project was inspired by “RASHT-A Partially Reconfigurable Architecture for Efficient Implementation of CNNs” published in IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 2022.

Developed a compiler using ANTLR and Python to automatically generate Python code for PyQt QLCDNumber widgets from a custom XML input. This tool streamlines the UI development process by translating a simple markup language directly into a functional GUI component.

Developed a desktop library management system using C# and WPF for the front-end and SQL Server for the back-end database. The application features distinct user roles (Administrator and Standard User), each with a unique dashboard and permissions tailored to their responsibilities.

Designed and implemented a VHDL simulation of a memory hierarchy featuring a multi-level cache. The system was driven by a simplified CPU model responsible for fetching and executing instructions, generating realistic memory access patterns to validate the cache’s performance.

Implemented and benchmarked multiple Constraint Satisfaction Problem (CSP) algorithms within a Sudoku solver. This project compares a standard backtracking approach against versions enhanced with Arc Consistency (AC-3) and the Minimum Remaining Values (MRV) heuristic. Each algorithm was evaluated on puzzles of varying difficulty by measuring execution time and the number of nodes expanded.

Developed a Tic-Tac-Toe AI agent using the Monte Carlo Tree Search (MCTS) algorithm. The agent builds a search tree over thousands of iterations; each iteration involves a process of UCB1-based selection, node expansion, random simulation, and backpropagation of results. The project includes a playable game loop for a human vs. AI match and a suite of unit tests to verify the agent’s logic.

Implemented a solution for the N-Queens problem using the Hill Climbing with Random Restart algorithm. The solver applies local search, starting from a random board state, to iteratively minimize the number of queen conflicts based on a heuristic function. A random-restart mechanism is integrated to escape local minima, with the solver’s performance benchmarked by its success rate and execution time over multiple trials.