Led a cross-functional initiative to develop and deploy a constraint-aware portrait cropping system, leveraging quadratic programming optimization integrated into production image workflows.
Built a deep learning CNN regression model to predict color editing parameters from yearbook images (100K+ dataset), achieving 3.5 mean LAB error and enabling automated high-volume color correction.
Deployed color correction model using OpenVINO-optimized ONNX inference and Triton Inference Server, with DVC-managed datasets and automated cross-runtime validation.
Researched self-supervised contrastive learning for latent color embeddings, reducing labeled training data requirements by 80%.
Migrated legacy GPU-based face landmark model to an optimized CPU ONNX pipeline; reduced infrastructure cost by 50% and increased throughput by 20%.
Architected a multi-stage GAN-based training pipeline for portrait hair matting and stray-hair removal; improved matte edge fidelity by 35% boundary IoU.
Engineered a large-scale synthetic data generation pipeline producing 5M+ labeled training samples, reducing manual annotation by 90% and improving model generalization.
Developed face super-resolution models using vector-quantized codebooks and transformers, trained on 6M+ images and deployed in a consumer mobile app, increasing user engagement by 30%.
Designed a lightweight residual-feature glare detection pipeline for eyeglass regions, reducing operation costs by 40%.
Built automated dataset validation tools using IoU/Dice consistency checks; removed 30% noisy samples and improved model convergence speed by 25%.
Developed a Python-based tool for testing and validation of ML models within deployment pipelines, improving reliability of evaluation workflows.
Built ensemble time-series forecasting models using LightGBM, optimizing hyperparameters with Optuna and analyzing feature importance with SHAP for supply chain demand prediction.
Research Assistant at Network and Multimedia Systems Lab (Prof. Mohamed Hefeeda).
Thesis: Proposed the first unsupervised single-image reflection removal framework using cross-coupled Perceptual Deep Image Priors. Published in IEEE TMM 2022. Achieved 37% improvement over prior methods and 24% PSNR gains.
Courses: Statistical Machine Learning, Deep Learning, Computational Photography, Design and Analysis of Algorithms.
Thesis: Deep Reinforcement Learning for Dynamic Reliability Aware NFV-Based Service Provisioning (Prof. Hamed Kebriaei, Prof. Vahid Shah-Mansouri). Published at IEEE GLOBECOM 2019.
Honors: Honored Alumni Award, ECE Department. Iran's National Elites Foundation.
Designed a Deep Reinforcement Learning (DQN) framework for reliability-aware resource allocation, optimizing service placement under bandwidth, compute, and reliability constraints.
Member of Iran's National Elites Foundation.
Recent presentations with the posters and reports
In this work, we aim to train a model for the task of visual question answering, using only a small number of labeled data.
Read more Read lessIn this work, we considered a dynamic reliability-aware NFV placement using DQN.
Read more Read less