Muhammad Umair

SQA Engineer & Tech Support

Auto Dataflow • Full time

Jun 2023 - Present • 3 yrs 4 mos

Performed comprehensive manual testing across multiple software releases, achieving high defect detection rates and maintaining superior product quality standards. • Designed and executed detailed test cases for functional, regression, and user acceptance testing (UAT), contributing to a measurable improvement in software reliability. • Collaborated cross-functionally with development teams to document, track, and resolve bugs using JIRA, streamlining the defect lifecycle and bug-fixing process. • Led integration of Selenium WebDriver automation testing framework to support and expand future test automation coverage. • Participated in all phases of the Software Development Life Cycle (SDLC), ensuring adherence to quality standards and timely delivery. • Improved software release cycle by reducing post-release defects by 25% through enhanced testing strategies and process optimization. • Simultaneously managed 30+ application testing projects, ensuring on-time delivery while maintaining strict adherence to QA protocols. • Optimized test coverage methodology, resulting in 20% faster turnaround for bug identification and resolution.

AI Projects

Fiverr • Freelance

Oct 2022 - Present • 4 yrs

AI vs. Human Text Classification System [Fiverr Delivery] Tech Stack: Python, TF-IDF, Logistic Regression, BERT, DistilBERT, Hugging Face Transformers, Scikit-learn, Google Colab • Designed and implemented an end-to-end NLP pipeline to classify text as AI-generated vs. human-written using both classical and transformer-based supervised learning approaches. • Trained and benchmarked two model tiers: TF-IDF + Logistic Regression (baseline) and fine-tuned DistilBERT/BERT (high-accuracy production model). • Evaluated models using accuracy, precision, recall, F1-score, confusion matrix, and ROC-AUC on strict train/validation/test splits with no data leakage or reshuffling. • Developed an API-ready inference function returning predictions with confidence scores and implemented model saving/loading for production deployment and future retraining. • Implemented long-text handling strategies, including truncation and chunking; delivered in Google Colab with full documentation, reproducibility controls, and reusable configuration. Process Mining & Anomaly Detection System [Fiverr Delivery] Tech Stack: Python, Pandas, NumPy, Scikit-learn, Jupyter, Process Mining (BPI Challenge Event Logs) • Designed and implemented a production-ready, end-to-end process mining and anomaly detection pipeline for eventlog data, converting a POC Jupyter notebook into a modular, reusable Python script. • Engineered features from raw activity logs — including case duration, event order patterns, rework frequency, and timing gaps — to detect anomalous process behavior across multiple process variants. • Trained and evaluated ML models on BPI Challenge-style event logs, validating performance and building a modular architecture (data ingestion, preprocessing, feature extraction, modeling, scoring) to support new log schemas and clients. • Implemented stress-testing strategies, including data perturbation and synthetic anomaly injection, to assess model generalization beyond training data.

MS Researcher

University of Engineering and Technology, Lahore • Lavoratore autonomo

Sep 2023 - Apr 2026 • 2 yrs 7 mos

SDLC Software Failure Prediction [MS Thesis] Tech Stack: Python, Scikit-learn, XGBoost, Random Forest, SVM, Logistic Regression, Pandas, Matplotlib, Seaborn, GridSearchCV • Conducting MS thesis research on predicting software failures from SDLC event data using supervised machine learning, directly applying domain expertise from 5+ years in QA. • Built a multi-model classification pipeline on a 4,500-record SDLC dataset, benchmarking Logistic Regression, SVM, Random Forest, and XGBoost across accuracy, F1-score, and ROC-AUC. • Applied GridSearchCV hyperparameter tuning on Random Forest (n_estimators, max_depth, min_samples_split) with 5- fold cross-validation to identify the optimal model configuration. • Performed feature importance analysis to identify the top SDLC factors contributing to software failures, providing actionable insights for QA and development teams. • Visualized model performance using confusion matrices, ROC curves, and feature importance charts; results inform real-world QA process improvements. Missing Child Identification Using Face Recognition & Age Progression [Personal Project] Tech Stack: Python, TensorFlow, Keras, ResNet50, GAN, OpenCV, NumPy, FG-NET Dataset • Built a deep learning system to identify missing children by matching facial features across different age groups using the FG-NET aging dataset and ResNet50-based face embeddings. • Designed and trained a Generative Adversarial Network (GAN) with a custom Generator (Conv2D + Transpose) and Discriminator architecture for realistic age progression of child facial images. • Generated 128-dimensional face embeddings via ResNet50 feature extraction; built a sequential age prediction network achieving low Mean Absolute Error (MAE) on test data. • Implemented a full image preprocessing pipeline, including normalization, resizing (128x128), and label parsing; saved trained models and embeddings for future inference and deployment.