Current Position
Master's student in AI at Ca' Foscari University of Venice
Graduation Date
Expected June 2027
Field of interest
Machine Learning · Computer Vision
I hold a Bachelor's degree in Computer Science from Iran and have four years of experience working as a backend developer. I am currently in my second year of a Master's in Artificial Intelligence at Ca' Foscari University of Venice, where I focus on machine learning and computer vision. I am seeking a backend developer position where I can apply my experience, or an internship as a Machine Learning Engineer starting from August 2026.
Machine Learning & AI
Computer Vision & Data
Programming & Frameworks
Tools & Development
NEAT · Self-Attention NEAT · Jul 2026 – in progress
KWIN is a cube-bodied character in a 3D room whose only brain is an evolved neural network — no supervised training, no gradient descent. It senses the world through depth rays and a first-person HSV frames feed, and NEAT evolves a CNN-based controller purely from a reward for eating green balls and avoiding red ones.
LION20 · Feb 2026
Benchmarked ten modeling approaches — from gradient-boosted trees to an attention-based CNN-BiLSTM — on a highly imbalanced, highly correlated ICP dataset. Missing values were handled with KNN imputation, classical-model features were reduced via PCA (95% variance retained), and class imbalance was addressed with SASMOTE plus class-weighted sampling. Best model (CNN-BiLSTM-Attention) reached 0.979 test AUC at 0.97 sensitivity / 0.90 specificity.
Model comparison (train/test metrics; test confusion matrix)
| # | Model | Tr AUC μ | Tr AUC σ | Tr t (s) | Te AUC | Sens | Spec | TP | TN | FP | FN | Te t (s) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | CNN-BiLSTM-Attention | 0.9992 | 0.0001 | 2237 | 0.9787 | 0.9696 | 0.8951 | 4626 | 128 | 15 | 145 | 3433 |
| 2 | Dilated-CNN | 0.9621 | 0.0009 | 96 | 0.9141 | 0.8405 | 0.8462 | 4010 | 121 | 22 | 761 | 182 |
| 3 | CatBoost | 0.9882 | 0.0009 | 1 | 0.9317 | 0.9382 | 0.8112 | 4476 | 116 | 27 | 295 | 2 |
| 1D-CNN (Base) | 0.9307 | 0.0106 | 75 | 0.8738 | 0.8132 | 0.7762 | 3880 | 111 | 32 | 891 | 143 | |
| XGBoost | 0.9845 | 0.0012 | 1 | 0.9203 | 0.9401 | 0.7692 | 4485 | 110 | 33 | 286 | 1 | |
| Attention-CNN | 0.9519 | 0.0043 | 188 | 0.8926 | 0.8742 | 0.7622 | 4171 | 109 | 34 | 600 | 630 | |
| Extra Trees | 0.9880 | 0.0011 | 1 | 0.9221 | 0.9554 | 0.7343 | 4558 | 105 | 38 | 213 | 1 | |
| LightGBM | 0.9983 | 0.0002 | 5 | 0.9467 | 0.9805 | 0.6993 | 4678 | 100 | 43 | 93 | 5 | |
| Random Forest | 0.9982 | 0.0003 | 12 | 0.9416 | 0.9782 | 0.6923 | 4667 | 99 | 44 | 104 | 15 | |
| CNN-LSTM | 0.9913 | 0.0005 | 495 | 0.9280 | 0.9602 | 0.6853 | 4581 | 98 | 45 | 190 | 947 | |
| ResNet-1D | 0.9759 | 0.0019 | 84 | 0.9019 | 0.9447 | 0.6643 | 4507 | 95 | 48 | 264 | 157 | |
| MLP | 0.9973 | 0.0002 | 49 | 0.9107 | 0.9799 | 0.6573 | 4675 | 94 | 49 | 96 | 105 |
3D Geometry and Computer Vision Course · Mar 2026
Built a 3D scanner from a single calibrated camera and a laser-line projector: with two reference planes at a known distance, the laser's intersection with each plane recovers per-pixel depth, and reprojecting across frames of a rotating object reconstructs a dense point cloud. Implemented the full pipeline — camera calibration, laser-plane detection, and depth reprojection — for the Geometric and 3D Computer Vision course.
Productivity project · Oct 2025
Anki is great for studying, but a browser-based version with AI built in is even more useful. Upload files in a few different formats and let an AI agent summarize the content into cards; decks are shareable and sync across mobile and desktop.
Introduction to Machine Learning · Sep 2024
Modified YOLOv8's architecture and training pipeline to beat the RePAIR project's published segmentation and detection benchmarks. Customized YOLOv8l reached 0.844 box mAP50 and 0.903 segmentation mAP50 — see the comparison below.
Customized YOLOv8l — validation metrics: bounding boxes (left) and instance segmentation (right).
| Model | Box P | Box R | Box mAP50 | Seg P | Seg R | Seg mAP50 |
|---|---|---|---|---|---|---|
| Customized YOLOv8l | 0.7866 | 0.8659 | 0.8439 | 0.8961 | 0.8113 | 0.9025 |
Productivity project · Oct 2025
An n8n and Gemini pipeline, backed by Postgres and served through FastAPI, that pulls, summarizes, and surfaces new CV/ML papers daily — built so I actually keep up with the literature instead of letting a reading list pile up.
Introduction to Artificial Intelligence · Apr 2023
Implemented Semi-Supervised SVM (S3VM) and a Newton-based Universum Twin SVM (Newton-UTSVM), then proposed an Unconstrained S3VM that blends both. Newton-UTSVM matched or beat S3VM's accuracy on 5 of 6 benchmark datasets while training over 12,000× faster (0.8s vs. 9,993s on the largest set).
Benchmark comparison: accuracy % (top) and runtime in seconds (bottom); bold = best in row for that metric.
| Dataset (samples × features) | Newton | S3VM constrained | S3VM unconstrained |
|---|---|---|---|
| Diabetes Pima (768 × 8) | 76.630.03 |
75.4813.50 |
75.100.49 |
| Ionosphere (350 × 34) | 86.550.03 |
86.5512.35 |
84.031.40 |
| Musk (476 × 166) | 86.420.25 |
81.48158.12 |
80.863.12 |
| Breast Cancer (568 × 31) | 97.410.03 |
97.9358.16 |
94.304.18 |
| Sonar (207 × 60) | 85.710.02 |
80.006.43 |
67.140.90 |
| Gender (5001 × 7) | 96.470.80 |
96.299993.28 |
95.7122.28 |
Jul 2020 · with Hossein Moosaei & Dave Musicant
NDC generates random centers for multivariate normal distributions, separating planes, and class labels; measures separability by points on the wrong side of the plane. Integer-valued for simplicity.
MC-NDCC generator · 3D scatter preview · CSV download
Loading chart…
Python, Django, Yolov5 · Feb 2023
A dentist-facing panel for managing patient records and X-rays, with a Yolov5 model flagging decayed or at-risk teeth for clinical review.
LION20 · Special Session 6, Digital Healthcare Systems · Accepted, Feb 2026
Accepted and presented at LION20's special session on digital healthcare systems. The paper tackles a highly imbalanced, highly correlated dataset for predicting intracranial pressure (ICP) signals — classical models, tree-based methods, and deep learning architectures, combined with feature engineering and data augmentation, to minimize false positives while keeping sensitivity, specificity, and AUC high.
Springer · Nov 2023
In this article we propose (NUTBSVM), a Newton-based approach for solving in the primal space the optimization problems related to Twin Bounded Support Vector Machines with Universum data (UTBSVM). In the NUTBSVM, the constrained programming problems of UTBSVM are converted into unconstrained optimization problems, and a generalization of Newton’s method for solving the unconstrained problems is introduced.
DeepLearning.AI, Stanford University, Prof. Andrew Ng · Oct 2025
ETS · Mar 2023
101/120
Ca' Foscari University of Venice
Master’s Degree in Artificial Intelligence and Data Engineering
Sep 2024 – exp. June 2027
| Course | CFU | Grade |
|---|---|---|
| Foundations of Artificial Intelligence and Machine Learning | 12 | 27/30 |
| Geometric and 3D Computer Vision | 6 | 30/30L |
| Image and Video Understanding | 6 | 27/30 |
| Software Architectures | 6 | 25/30 |
| Algorithms and Learning over Massive Data | 12 | 22/30 |
| Calculus and Optimization | 6 | 23/30 |
| Information Retrieval and Web Search | 6 | 24/30 |
| Statistical Inference and Learning | 6 | 26/30 |
| Applied Probability for Computer Science | 6 | 28/30 |
| Advanced Data Management | 6 | 26/30 |
| Cloud Computing and Distributed Systems | 6 | September 2026 |
| Cryptography | 6 | September 2026 |
| Deep Learning for Natural Language Processing | 6 | Fall 2026 |
| Final thesis | 24 | Fall 2026 |
| Internship | 6 | Fall 2026 |
University of Bojnord
Bachelor’s Degree in Computer Science
Sep 2012 – Jun 2016
| Course | Grade |
|---|---|
| Fundamentals of Computer and Programming | 18.5/20 |
| Algorithms and Data Structures | 20/20 |
| Advanced Programming | 16.52/20 |
| Differential Equations | 20/20 |
| Fundamentals of Combinations | 20/20 |
| General Mathematics | 20/20 |
| Computer Graphics | 19/20 |
| Principles of Operating Systems | 20/20 |
| Theory of Calculation | 18/20 |
| Topics in Computers Sciences | 20/20 |
| Principles of Software Design | 20/20 |
| Bachelor Project of Computer Sciences | 20/20 |
Aug 2022 – Mar 2024
Respina Network & Beyond, Tehran, Iran
Respina is one of Iran's larger telecom providers — dedicated internet access, SIP-Trunk, hosted PBX, data center colocation. I joined the Hosted-PBX (Nexfon) team and spent most of my time on two problems: billing and call capacity. I rebuilt the billing pipeline on CGRATES for real-time charging, fixing a long-standing accuracy issue, and rewrote our monthly reporting queries — parallelized with Celery — which cut a 30-minute job down to 40 seconds. On the call-handling side, I refactored our Asterisk-ARI integration into an event-driven Flask microservice, containerized it with Docker, and added multi-processing; concurrent call capacity per instance went from 25 to 130, roughly a 5× drop in infrastructure load. I also set up monitoring with Prometheus, Loki, and Grafana, and worked with the DevOps team on migrating deployments to Kubernetes.
Nov 2021 – Jul 2022
ANIL Web design studio, Tehran, Iran
Worked alongside a small, sharp team on two production sites — RadmanPack and PelikanIran — building the Laravel backends and the REST APIs a React frontend talked to. Handled authentication, API design, and Docker-based deployment on Linux. It was my first real exposure to structured code review and Agile process.
Oct 2020 – Jul 2021
AvinAvisa Lab, Tehran, Iran
Joined the Blockchain Lab at Amirkabir University of Technology to work on Polychain, a peer-to-peer crypto exchange. My main contribution was the trade-matching engine — an algorithm inspired by the Knapsack problem that selected trade pairs based on volume, VIP tier, and a few other dynamic factors. I also built the Ethereum and TRON integrations, so a good part of my time went into making sure on-chain transactions stayed both fast and correct. Small team, MERN stack, a lot of learning by doing.
Jul 2018 – Jul 2020
Army, Tehran, Iran
Military service is mandatory in Iran and lasts two years — without it, you can't get a passport to leave the country. I used the time productively: co-authored a publication and a research project with Dr. Hossein Moosaei (Charles University) and Dr. David Musicant (Carleton College), and taught myself machine learning and image processing on the side.
Aug 2016 – Apr 2018
Ishaya, Tehran, Iran
My first job. Joined Ishaya as an Android intern and stayed on full-time, working on Ponila — the first Persian-language content recommendation app. I built the Android client and worked with the backend and data science teams to hook it up to a recommendation engine driven by semantic and syntactic analysis of Persian text. A good crash course in shipping a real product end to end.
