About me
Machine Learning Engineer, specializing in Computer Vision with more than 3 years of practical experience in the industry and academic world. Building and maintaining real-time intelligent systems in a variety of challenges from face authorization and vision-based recommendation to multi-modal video indexing systems. Have worked in agile environments with teammates from 4 to 20 colleagues. Equipped with back and front-end technologies and hands-on experience for making production-level services for AI-based modules. Actively working on open-source projects on my GitHub. A mentor, buddy, and friend for newcomers in the workplace and programming institutions.
Work Experience
Missions:
- Experimented with different transformer-based text encoders in lora training of stable diffusion models
- Designed a background removal and replacement system based on stable diffusion models
- Trained several models based on textual inversion method for defect detection in images
- Built a fast visualization and exploration system for large imagery databases based on similarity indexing
- Improved the pipeline’s accuracy by approximately 6%, reaching above 97% on well-known NSFW datasets in the literature
- Optimized the pipeline's backbone to have at least 50% smaller with increased performance.
- Refactored a massive amount of code and converting it to a clean reusable core module for deployment
- Improved the overall precision and recall of the system by a margin of nearly 12% and 14% on 500 people
- Improved the overall precision and recall of the system by a margin of nearly 7% and 10% on a in-house dataset including 1600 iranian celebrities
- Implemented a BERT-based language model with a downstream task of spell checking on the Persian language
- Designed an OOP-based interface for it to be deployable in any intellisense system.
- Reached WER of lower than 3.5% on a thorough dataset of 20 million Persian sentences, outperforming the SOTA in the literature
- Mentored and coached a teammate for PyTorch library in several sessions
- Collected over 10k samples and making an automated pipeline of labels for localization and recognition
- Improved its accuracy more by more than 7% on cursive fonts by integrating Attention module
- Several improvements on the speed and performance of localization module
- Design a self-supervised approach for learning discriminative features by using contrastive learning and masked encoding
- Improved human-voted relevancy factor, meaning the 20 first recommendations’ similarity to the query based on 10 people
- Decreasing time load to 0.67 of the common implementations
- Automatic and efficient translation of the benchmark datasets using scraping technologies
- Trained a Seq2Seq model for translating different languages to Persian; this module was made to provide the dataset for train the image captioner
- Implemented a novel Seq2Seq-based supervised approach for detecting the most important scenes of soccer match videos
- Implemented a module for retrieving iranian celebrities face images for video-indexing purposes
- Created several sound preprocessing modules, including human voice detector, sound diarization, and sound noise removal
- Training and maintaining Wav2Vec and DeepSpeech for Persian language
- Achieved 1.43 WER on “common voice” benchmark
- Build a product recommendation system, integrating a visual features and meta specifications
- Integrated Postgresql database and FAISS-based similarity search pipeline for the products
- Providing a clean interface for the deployment
- Developed a cross-platform software using PyQt for a private hospital for recording medical data from Traumatic brain injury (TBI)
- Equipped with features for training and evaluating different supervised machine learning-algorithms with a user-friendly UI
- Implementing SQL database and the required connections
- Testing and Maintaining
- Designed a document reader for a law firm for automating customer services
- Paragraph segmentation, signature extraction, OCR
- Implemented a user-friendly GUI-based app for cataract detection for a private medical services center
- Utilized light-weight deep models for better performance with low cost
- Achieved a trustworthy performance of 98.87 TPR on local dataset
- Implemented more than 4 in-house applications for faciliation of engineering procedures
- BIM Mdeller and coordinator in more than 6 structural projects
Projects
Skills
Python
Pytorch
Tensorflow
Numpy
Matplotlib
Seaborn
Pandas
JavaScript
Django
Flask
FastAPI
React
SQL
HTML
CSS
Bootstrap
Github
Linux
Publications
[1] Soroush Hashemifar, Abdolreza Marefat, and Javad Hassannataj Joloudari, “FRA: A novel Face Representation Augmentation algorithm for face recognition ”, 2022
(Expert Systems with Applications)
Show publication[2] by Javad Hassannataj Joloudari, Abdolreza Marefat ,Mohammad Ali Nematollahi,Solomon Sunday Oyelere 4, and Sadiq Hussain, “Effective Class-Imbalance learning based on SMOTE and Convolutional Neural Networks”, 2023
(MDPI - Applied Sciences)
Show publication[3] Abdolreza Marefat, Hassannataj Joloudari, Maryam Rastgarpour, “A Transformer-Based Algorithm for Automatically Diagnosing Malaria Parasite in Thin Blood Smear Images Using MobileViT”, 2022
(Submitted)