Onisa Mapunda

ONISA PASCAL Dar es Salaam, Tanzania - - SUMMARY Machine Learning Engineer with 5+ years of experience designing and implementing production-grade ML/Al solutions across computer vision, NLP, and IoT domains. Demonstrated expertise in the full ML development lifecycle from data collection to model deployment, with a strong track record of delivering systems that improve efficiency by 15-40%. Skilled in deep learning architectures, unsupervised learning, and developing custom algorithms for specialized applications in healthcare, transport and agricultural sectors. SKILLS Python, SQL, C++ PyTorch, TensorFlow, Keras, scikit-learn, Hugging Face Docker, Kubernetes, CI/CD, MLflow, Weights & Biases, DVC AWS (SageMaker, Lambda, EC2), GCP (Vertex AI), Azure ML Spark, Hadoop, Kafka, Airflow PostgreSQL, MongoDB, Redis, Elasticsearch Computer Vision, NLP, Time Series Analysis, Anomaly Detection, Unsupervised Learning Raspberry Pi, IoT Sensors, Edge Computing EXPERIENCE ML ENGINEER Counterest April 2024 - Present, Barcelona, Spain • Optimized computer vision algorithms for crowd density counting in surveillance systems, reducing Mean Absolute Error by 22% while maintaining real-time performance. • Implemented transfer learning techniques to adapt models for varying environmental conditions, increasing accuracy in low-light scenarios by 30%. • Designed and deployed edge computing solutions for distributed processing, reducing cloud computing costs by 35%. ML ENGINEER Nolla Africa January 2022 - August 2024, Mbeya, Tanzania • Designed and implemented ML systems for wildlife detection and classification with 92% accuracy, deployed across web platforms and field monitoring stations. • Diagnosed and repaired under performing machine learning pipelines, improving operational efficiency by 40% through code optimization and architectural improvements. • Partnered with cross-functional teams to integrate ML models into production systems, improving prediction accuracy by 15% over 6 months. • Developed automated maintenance and monitoring strategies for critical ML systems, reducing downtime by 25% and enhancing system reliability. ML RESEARCH ASSISTANT Mbeya University November 2022 - June 2023, Mbeya, Tanzania • Conducted customer behavior analysis using Python, NoSQL, and advanced statistical methods, uncovering key patterns that led to 18% improvement in prediction models. • Developed and implemented LSTM Neural Networks for time-series prediction, achieving 87% accuracy on forecasting tasks. • Enhanced research capabilities through machine learning libraries including scikit-learn, NumPy, Pandas, and visualization tools such as Matplotlib and Seaborn. • Created custom data preprocessing pipelines that reduced training time by 30% while maintaining model performance quality. PUBLICATIONS Unsupervised Power Quality Events Clustering with Transformer-Based Autoencoders Manuscript submitted • IEEE Transactions on Power Systems • Proposed an unsupervised deep learning model using transformer autoencoders to cluster power quality disturbances (sags, swells, transients) from unlabeled voltage waveform data. • Achieved 0.62 silhouette score with 8 distinct clusters via K-means on latent features, validated by attention heatmaps and t-SNE visualization. • Demonstrated the model’s ability to prioritize critical temporal features (e.g., interruptions, harmonics) via self-attention mechanisms, reducing manual labeling needs for grid diagnostics. PROJECT IoT Vital Check-Up System for Patient Monitoring • Architected and deployed an IoT ecosystem using Raspberry Pi to collect and process vital signs data from bed-mounted sensors. • Developed real-time analytics pipeline processing 500+ data points per minute with 99.9% uptime. • Implemented predictive algorithms detecting potential health deterioration 40 minutes before critical events (average). • Created secure remote monitoring interface reducing nurse response time by 35% and improving patient outcomes. Plant Disease Detection and Heat Map Visualization System • Created and fine-tuned a convolutional neural network for detecting 38 different plant diseases with 94% precision. • Implemented gradient-weighted class activation mapping (Grad-CAM) to generate heat maps highlighting disease-affected areas. • Developed anomaly detection module ensuring 97% accuracy in verifying whether images belong to target plant species. • Built a mobile-friendly interface allowing farmers to access diagnostic results in remote areas with limited connectivity. Facial Security and Data Analysis System for Hospital Management • Implemented facial recognition algorithms achieving 96% identification accuracy for secure patient identification and access control. • Developed predictive algorithms for chronic disease detection using patient data and historical records, improving early diagnosis rates by 25%. • Applied deep learning techniques to analyze CCTV images, increasing accuracy of detecting abnormal bleeding by 30%. • Integrated the system with hospital management software, optimizing bed allocation and resource management, resulting in 15% reduction in administrative overhead. Automated PDF Organization and Metadata Extraction • Developed an automated system to download, rename, and organize PDF files from websites with 99% reliability. • Integrated Al components using OpenAI API and OCR technology to fill missing metadata, improving accuracy by 40%. • Implemented custom SKU matching system achieving 95% accuracy in resolving file naming discrepancies. • Reduced manual document processing time by 75%, significantly improving organizational workflow. • Proposed an unsupervised deep learning model using transformer autoencoders to cluster power quality disturbances (sags, swells, transients) from unlabeled voltage waveform data. • Achieved 0.62 silhouette score with 8 distinct clusters via K-means on latent features, validated by attention heat-maps and t-SNE visualization. • Demonstrated the model’s ability to prioritize critical temporal features (e.g., interruptions, harmonics) via self-attention mechanisms, reducing manual labeling needs for grid diagnostics. EDUCATION DIPLOMA IN MECHATRONICS AND ROBOTICS AUTOMATION Mbeya University • Mbeya, Tanzania • November 2020 - July 2023 • Relevant coursework: Machine Learning Fundamentals, Computer Vision, Embedded • Systems, Robotics Control Systems, Sensor Networks • Capstone Project: Automated Precision Agriculture System using Computer Vision and loT.