CV

Research Interests

Deep Learning · Deep Reinforcement Learning · Medical AI · Efficient Deep Learning · Time Series Forecasting

Education

Shahid Beheshti University | Tehran, Iran
Master’s Degree, Data Science | Sep 2024 – Sep 2026

• Thesis: Designing a Dynamic Forecasting and Risk Management Framework in Stock Trading: A Probabilistic Deep Learning Approach (Supervisor: Dr. Ehsan Bahrami Samani)

Iran University of Science and Technology | Tehran, Iran
Bachelor’s Degree, Computer Engineering | Sep 2019 – Nov 2023

• Thesis: Storing Big Data on DNA: Design and Implementation of a Secure Pipeline Using Cryptographic Protocols (Supervisor: Dr. Marzieh Malekimajd)

Publications & Preprints

1. Kazemi, S. M., Azarpour, A., Riahi Madvar, M. CRNN-KAN: A Lightweight Hybrid Framework for Robust Phonocardiogram Classification. Informatics in Medicine Unlocked. Status: UNDER REVISION.
2. Kazemi, S. M., Heydari, P. RL-WavKAN: An Ultra-Lightweight Reinforcement-Learned Wavelet Kolmogorov–Arnold Network for Phonocardiogram Classification. Computer in Medicine and Biology. Status: Under Review.
3. Azarpour, A, Kazemi, S. M. SW-KAN: Kolmogorov–Arnold Networks with Stieltjes–Wigert q-Orthogonal Polynomials. Computers and Mathematics with Applications. Status: Under Review.
4. Riahi Madvar, M., Kazemi, S. M., Yousefi, S. Scalable unsupervised outlier detection in high-dimensional big data via collision-based subspace selection and deep reinforcement learning. Applied Soft Computing. Status: Under Review.
5. Kazemi, S. M., Bahrami Samani, E. Probabilistic Joint Forecasting of Bivariate Time Series with LSTM via Dynamic Student-$t$ Likelihood. Computational Economics. Status: UNDER REVISION.
6. Kazemi, S. M., Ghanbary, S., Riahi-Madvar, M. Discovering Actionable Marketing Insights: Rare Association Rule Mining for Customer Behavior Analysis. 2025 10th International Congress on Fuzzy and Intelligent Systems (CFIS). DOI: 10.1109/CFIS68949.2025.11652160.

Research Experience

Research Assistant

Shahid Beheshti University, Tehran, Iran | Supervisor: Dr. Mahboobeh Riahi-Madvar | Apr 2025 – Sep 2026

• Lightweight Hybrid Deep Learning for Medical Time Series (CRNN-KAN) (Jan 2026 – Sep 2026)
  • Proposed a novel, parameter-efficient CRNN-KAN architecture for robust phonocardiogram (PCG) classification.
  • Integrated a CNN-BiLSTM backbone for spatiotemporal feature extraction with an EfficientKAN classification head, utilizing SiLU and sinusoidal basis functions.
  • Achieved competitive AUC scores across three clinical datasets (PhysioNet, CirCor, PASCAL) while reducing trainable parameters to 1.31M (a 100x reduction vs. VGG16).
  • Validated the physiological interpretability of the learned acoustic representations using Score-CAM visualizations.
• Scalable Outlier Detection in Big Data (mCRMR-GPPO) (Apr 2025 – Dec 2025)
  • Developed a distributed, unsupervised anomaly detection framework for high-dimensional big data, deployed on Apache Spark using the MapReduce programming model.
  • Formulated a novel subspace selection criterion (mCRMR) leveraging Locality-Sensitive Hashing (LSH) and collision-based redundancy to isolate informative features.
  • Designed a hierarchical Gated Proximal Policy Optimization (GPPO) architecture, filtering normal instances efficiently and dedicating deep neural network processing strictly to complex anomaly patterns.
  • Demonstrated near-linear computational scalability across cluster nodes and achieved near-perfect AUC ($\approx$99%) on synthetic and real-world high-dimensional datasets.
• Rare Association Rule Mining for Bank Marketing (Jul 2025 – Sep 2025)
  • Designed an interpretable data mining pipeline to replace black-box marketing models, extracting actionable “flip rules” to explain customer behavior shifts.
  • Implemented an adaptive dynamic minimum-support mechanism and MinHash LSH to mitigate the rule explosion problem and efficiently discover rare, high-impact patterns.
  • Identified precise decision boundaries that successfully improved targeted marketing campaign conversion rates from 11.7% to 64.7% (a 5.53x lift).

Research Assistant

Shahid Beheshti University, Tehran, Iran | Supervisor: Dr. Ehsan Bahrami Samani | Oct 2025 – Dec 2025

• Probabilistic Joint Forecasting of Bivariate Time Series
  • Engineered a Bivariate Probabilistic LSTM framework to jointly forecast returns, volatility, and dynamic cross-asset correlations for commodity markets (Gold and Brent Crude Oil).
  • Parameterized a dynamic Student-$t$ likelihood optimized via Negative Log-Likelihood (NLL) to explicitly model heavy-tailed distributions and extreme financial shocks.
  • Produced well-calibrated 95% predictive intervals and outperformed SOTA baselines (Transformer, TCN, N-BEATS) in accuracy and tail risk quantification.

Independent Researcher

Collaborative Research, Tehran, Iran | Jan 2026 – Present

• 1D Reinforcement-Learned Wavelet KAN (RL-WavKAN) (Jan 2026 – Present)
  • Developed an ultra-lightweight 1D Wavelet Kolmogorov-Arnold Network for direct raw audio classification, bypassing computationally heavy 2D spectrogram transformations.
  • Formulated a custom Proximal Policy Optimization (PPO) agent to dynamically optimize the scale and translation of continuous wavelet basis functions during training.
  • Reduced the computational footprint to 145K parameters and $\sim$0.0003 GFLOPs, achieving an inference latency of 0.036 ms/sample for real-time cardiovascular screening.
• Stieltjes-Wigert $q$-Orthogonal Polynomials in KANs (SW-KAN) (Jan 2026 – Sep 2026)
  • Developed the SW-KAN architecture, replacing standard B-spline activations with Stieltjes-Wigert $q$-orthogonal polynomials to improve parameter efficiency and gradient stability.
  • Resolved the domain mismatch problem by introducing a smooth exponential-of-tanh domain mapping from $\mathbb{R}$ to the semi-infinite support $(0, \infty)$.
  • Leveraged a numerically stable three-term recurrence for $\mathcal{O}(N)$ evaluation, outperforming 18 baseline polynomial KAN variants on image classification and 2D function approximation tasks.

Undergraduate Research Assistant

Iran University of Science and Technology, Tehran, Iran | Supervisor: Dr. Marzieh Malekimajd | Sep 2022 – Sep 2023

• Secure Big Data Storage on DNA
  • Investigated DNA as an ultra-high-density, long-term storage medium for Big Data, addressing capacity and longevity limitations in conventional storage architectures.
  • Developed a complete data-to-DNA simulation pipeline in Python, successfully integrating file-to-binary conversion, Huffman coding for lossless data compression, and DNA sequence mapping.
  • Enhanced the standard DNACloud software framework by introducing a robust cryptographic security layer to protect synthesized genetic data.
  • Implemented and benchmarked AES and OTP encryption algorithms, evaluating computational efficiency and execution time across variable data payload sizes (up to 10 million characters).

Teaching & Academic Experience

• Teaching Assistant: Machine Learning, Multivariate Analysis (Instructor: Dr. Sakineh Dehghan)
• Teaching Assistant: Machine Learning, Advanced Machine Learning, Pattern Recognition (Instructor: Dr. Ahmad Ali Abin)
• Head TA & TA: Advanced Data Mining (Head TA), Machine Learning (Instructor: Dr. Mahboobeh Riahi-Madvar)
• Teaching Assistant: Big Data (Instructor: Dr. Hamed Malek)
• Teaching Assistant: Deep Learning (Instructor: Dr. Zeinab Hajimohammadi)

Selected Technical Projects

• Multimodal Vision-Language Medical Report Generator (CXR-Report-Generator): Developed an end-to-end encoder-decoder pipeline combining Vision Transformer (ViT) and GPT-2 to automatically generate clinical reports from Chest X-ray images via cross-attention mechanisms. Tools: PyTorch, Hugging Face Transformers, Python, OpenCV. [GitHub]

• FracAtlas Bone Fracture CAD Pipeline: Built a multi-task computer-aided diagnosis system utilizing DenseNet-121 for global classification and ResNet34-UNet for precise semantic segmentation, integrated with Meta’s SAM for zero-shot boundary delineation. Tools: PyTorch, Segmentation Models, OpenCV, Python, Segment Anything (SAM). [GitHub]

• DeepRC: Immune Repertoire Classification: Implemented a deep Multiple Instance Learning (MIL) framework using 1D CNN and LSTM sequence encoders paired with attention pooling to predict disease status from high-throughput immunosequencing data. Tools: PyTorch, SciPy, Scikit-Learn, H5py, Python. [GitHub]

• Multimodal Semantic Image Search & Recommender: Architected an end-to-end visual search engine leveraging OpenAI CLIP for deep semantic feature extraction and Qdrant vector database for high-performance K-NN retrieval, backed by FastAPI and React. Tools: FastAPI, React.js, Qdrant, CLIP, PyTorch, Python. [GitHub]

• xv6 OS Kernel Enhancements: Modified MIT’s xv6 operating system kernel by implementing custom system calls, dynamic process accounting, and advanced Multi-Level Queue (MLQ) CPU scheduling architectures. Tools: C, x86 Assembly, QEMU, Linux. [GitHub]

• Knowledge Engineering & Semantic Reasoning Suite: Developed macroeconomic ontologies in Neo4j/OWL and rigorously benchmarked GPT-4.1 multi-hop reasoning capabilities over structured biomedical and scientific knowledge graphs. Tools: Neo4j, Cypher, RDFLib, NetworkX, GPT-4.1, Python. [GitHub]

• OpenUnderstand Static Code Analyzer: Engineered an open-source, Python-based static code analysis framework utilizing ANTLR4 AST parsing for Java to extract complex software metrics and reverse-engineer architectural dependencies. Tools: Python, ANTLR4, Abstract Syntax Trees (AST). [GitHub]

• Intelligent Multi-Tool Movie Recommender: Built a hybrid recommendation engine integrating TF-IDF content filtering, Deep Neural Collaborative Filtering (NCF), and an autonomous LLM routing agent for natural language text-to-SQL factual retrieval. Tools: TensorFlow, Keras, Scikit-Surprise, Groq API, SQLite, Streamlit. [GitHub]

Technical Skills

• Programming Languages: Python, C/C++, SQL, R, MATLAB, Bash.
• AI & Deep Learning Libraries: PyTorch (Geometric, Torchaudio, Torchvision), TensorFlow, Hugging Face, XGBoost, Scikit-Learn, OpenCV, CUDA.
• Architectures & Paradigms: Sequence Models (KAN, LSTM, TCN), Foundation Models (ViT, CLIP, SAM), Deep RL (PPO, DQN, Contextual Bandits), Probabilistic Modeling, LLM Agents & RAG.
• Data Engineering & MLOps: PySpark, Vector DBs (Qdrant, ChromaDB), Relational DBs (PostgreSQL), Docker, FastAPI, Linux.
• Developer Tools: Git, LaTeX, GNU Make.

Selected Coursework

• Machine Learning & AI: Machine Learning (19.75/20), Deep Learning (19.5/20), Advanced Data Mining (19.25/20), Multi-Agent Systems [Focus: RL & LLMs] (19/20), Artificial Neural Networks (18.5/20)
• Mathematics & Foundations: Differential Equations (19/20), Multivariate Statistical Analysis (18.5/20), Engineering Probability & Statistics (17/20)
• Data Systems & Analytics: Big Data Analytics (17.25/20)

Honors & Awards

• Top 1.5% Nationwide: Ranked in the top 1.5% among ~15,000 participants in the Iranian Nationwide M.Sc. Entrance Exam

• Top 0.6% Nationwide: Ranked in the top 0.6% among >160,000 participants in the Iranian Nationwide B.Sc. Entrance Exam