Technologies: Python, ROS (Robot Operating System), PyTorch, ANNOY, 3D Sensors, Real-time Processing

Description: Developed an end-to-end framework for predicting human hand motion in shared workspaces to prevent robot-human collisions. Created custom dataset by simulating collision scenarios and collecting 3D trajectory data. Implemented 3D cell quantization for spatial discretization and developed vector embedding models for motion prediction. Utilized ANNOY (Approximate Nearest Neighbors Oh Yeah) for efficient nearest neighbor search enabling real-time predictions. System achieved superior accuracy compared to baseline methods while maintaining sub-100ms latency requirements.

Impact: Enhanced safety in collaborative robotics environments, enabling faster and safer human-robot interaction workflows.

Technologies: Python, PCL (Point Cloud Library), OpenCV, Scikit-learn, KITTI Dataset, ROS/RVIZ

Description: Designed and implemented 3D-PSH (3D Point Spatial Histogram), a lightweight object detection framework optimized for resource-constrained devices. Developed adaptive clustering algorithms for point cloud segmentation, computed 3D spatial histograms as feature descriptors, and applied Bag of Visual Words (BoVW) approach with SVM for classification. System was validated on KITTI dataset and real-world LiDAR sensors, achieving competitive accuracy with 3-5x faster inference compared to deep learning approaches.

Impact: Enabled real-time 3D object detection on edge devices without GPU requirements, suitable for autonomous vehicles and robotics applications.

Technologies: Python, Flask/Django, OpenCV, Face Recognition APIs, MySQL, JavaScript, IP Camera Integration

Description: Architected and deployed a production-grade smart surveillance solution for government premises. Integrated multiple IP cameras into a unified system performing real-time face detection, recognition, and tracking. Built RESTful APIs for camera management and data retrieval. Developed real-time web dashboard for monitoring and analytics. Implemented database schemas for efficient storage and querying of recognition events. System handles 10+ concurrent camera streams with sub-second recognition latency.

Impact: Automated visitor tracking reducing manual check-in time by 80%, enhanced security with real-time alerts for unauthorized access.

Technologies: Python, TensorFlow/Keras, Pandas, NumPy, Scikit-learn, Data Augmentation Techniques

Description: Developed deep neural network models for classifying biomedical samples based on Raman spectroscopic signatures. Implemented data preprocessing pipeline handling noisy spectral data including normalization, baseline correction, and augmentation. Designed CNN and fully-connected architectures optimized for 1D spectral data. Performed extensive hyperparameter tuning and cross-validation achieving 95%+ classification accuracy across multiple biomedical classes.

Impact: Automated biomedical sample classification reducing analysis time from hours to seconds, enabling high-throughput screening applications.

Technologies: Python, PyTorch, Transformers (BERT), LSTM, Flask, PostgreSQL, Data Scraping Tools

Description: Built end-to-end sentiment analysis system for multilingual social media content (English, Urdu, Roman Urdu). Created benchmark dataset for Roman Urdu (low-resource language) through web scraping and manual annotation. Developed custom tokenizer and embedding layer for Roman Urdu text. Implemented attention-based BiLSTM and fine-tuned multilingual BERT models. Built REST API for real-time sentiment prediction and batch processing. Deployed system analyzing 100K+ tweets daily during COVID-19 for public opinion monitoring.

Impact: Enabled real-time public sentiment tracking for government decision-making during critical periods, achieved F1-score of 0.89 on Roman Urdu dataset.

Technologies: Python, OpenAI API, Hugging Face Transformers, GPT models, Prompt Engineering, Statistical Analysis

Description: Developed automated system for identifying persuasive rhetorical strategies in crowdfunding campaigns using large language models. Designed prompt engineering strategies for zero-shot and few-shot classification. Compared multiple LLM architectures (GPT-3.5, GPT-4, BERT variants) against ground truth annotations. Performed correlation analysis between rhetorical strategies and funding success metrics. Built data pipeline processing 10K+ campaign texts with statistical validation.

Impact: Provided actionable insights for campaign optimization, identified key rhetorical elements correlated with 40% higher funding success rates.