• Teaching Assistant: Web Programming (2022–2023): For 2 semesters, Supported BS (7th semester) students, Provided coding, debugging, and MERN Stack (MongoDB, Express.js, React.js, Node.js) support over two semesters.

• Final Year Project: Intelligent Waste Management System - Developed a custom-trained Mask R-CNN model for garbage detection with 94.34% accuracy, implemented volume measurement techniques, and deployed using Django and REST APIs.

• Plant Species Identification with AI: Conducted research on advanced plant species identification by integrating Vision Transformer (ViT) models with metadata fusion, achieving 97.27% classification accuracy and 0.9842 Mean Reciprocal Rank.
• Multi-Modal Deep Learning Architecture: Developed a hybrid approach combining deep visual feature extraction with geographical and taxonomic metadata to enhance identification of morphologically similar species, significantly outperforming traditional CNN-based methods.
• Data Preprocessing and Analysis: Implemented comprehensive preprocessing pipeline with data augmentation techniques to improve model robustness across diverse environmental conditions, utilizing PyTorch for model development and evaluation.

• Intelligent Waste Management System: Developed a custom-trained Mask R-CNN model for detecting garbage in open areas and dustbins, achieving 94.34% accuracy for each waste class, enhancing environmental monitoring capabilities.
• Dataset Creation and Labeling: Created and meticulously labeled a custom dataset of approximately 500 images across 3 waste classes, establishing the foundation for accurate model training and validation.
• Volume Measurement Techniques: Implemented advanced computer vision algorithms for garbage volume measurement, enabling precise segmentation of waste areas to optimize collection schedules and resource allocation.
• Web Application Development: Deployed the trained model using Django framework and REST APIs, ensuring a robust and scalable application for real-time waste monitoring and management.

• Digital Ink Recognition R&D: Conducted research in digital ink recognition with focus on handwriting generation using reference and target text. Trained and evaluated custom models to mimic user handwriting.
• Model Development and Accuracy Improvement: Designed and trained RNN, LSTM-based, Transformer-based, and Diffusion-based models on IAMonDB and Brush Online datasets. Achieved an accuracy improvement from 89% to 93% in handwriting generation.
• Semantic Similarity Search with Vector DBs: Built a notebook content search system using ChromaDB and OpenAI LLMs to enable context-aware retrieval through embedding-based similarity.
• System Integration and Backend Engineering: Integrated model APIs using FastAPI and RESTful interfaces. Supported deployment and service integration with ReactJS frontend and Kotlin-based components.
• MLOps and Infrastructure: Containerized services with Docker and utilized PyTorch and TensorFlow for training and serving models in production.
• Startup Contribution: Worked closely with the startup founded by Rich Miner (Co-founder of Android), contributing to R&D.

• LLM Fine-Tuning and QA Systems: Fine-tuned LLaMA 3.2 and other foundation models (DeepSeek, Qwen2.5, GPT) using the QLoRA method on custom PDF datasets, building domain-specific question-answering systems.
• RAG System with Performance Boost: Developed a personalized RAG application using LangChain, Hugging Face, Qdrant, and Chrome integration. Improved answer accuracy from 70% to 80%+ through enhanced context retrieval techniques.
• Offline LLM Deployment and Inference: Deployed 70B GGUF LLMs locally using llama.cpp on A100 GPUs without internet. Integrated Open WebUI to enable RAG-based chatbot functionality in disconnected environments.
• AI Pipelines with Graph & Vector DBs: Engineered hybrid pipelines using LangChain and LangGraph, combining vector and graph databases for efficient, structured knowledge retrieval.
• Time Series Forecasting for Defense Systems: Created a synthetic dataset of 1 million records to train LSTM models for forecasting and anomaly detection. Boosted accuracy from 75% to 89%.
• Explainable AI in Anomaly Detection: Built Random Forest models with SHAP-based explainability to highlight key threat indicators for air defense systems.
• Computer Vision with YOLO: Trained YOLOv8 and YOLOv9 models on custom datasets and deployed them via Flask-based web apps for real-time object detection.
• Geospatial AI for Flight Pattern Simulation: Used QGIS and ML to analyze location-based data and simulate realistic flight paths using a hybrid of rule-based and machine learning systems.

• NLP Solutions: Implemented NLP solutions utilizing LLMs, LayoutLMv2, LayoutLMv3, LiLTv2, OCR, and GPT-2 technologies.
• Document Extraction Systems: Developed chatbots and document extraction systems using BERT-based models and LangChain.
• Proof of Concept (POC) Development: Created POC applications with Flask demonstrating NLP capabilities using OpenAI APIs.
• Model Optimization: Applied RAG techniques to LLaMA2 and LLaMA3 models, optimized models using LORA and QLORA methods.
• LLM Fine-tuning: Fine-tuned LLMs for customized application needs and enhanced performance.
• Team Mentoring: Mentored and guided junior engineers and interns, fostering professional development in AI and NLP.

• Data Preparation: Collected, labeled, and preprocessed fish datasets for classification, segmentation, and object detection.
• Fish BMI and Monitoring Systems: Calculated fish BMI using image analysis and implemented paddle wheel detection for monitoring through stereo-vision techniques.
• PDF-based NLP Retrieval System: Developed a query-based search system across uploaded PDFs to improve data analysis and retrieval efficiency.

• Research Paper: A novel approach integrating Vision Transformer (ViT) models with metadata fusion to enhance plant species classification accuracy to 97.27%, addressing challenges in distinguishing morphologically similar species.
• Key Contributions: Developed an AI-driven methodology combining deep visual feature extraction with geographical and taxonomic metadata, demonstrating significant improvements over traditional CNN-based approaches.

• Recognition: Honored as the first student team from University of Kotli (AJK) to present our Intelligent Waste Management System at this prestigious event, receiving a participation shield for demonstrating innovative application of computer vision techniques.

• Technical Instruction: Developed and delivered comprehensive 3-month curriculum on Machine Learning and Deep Learning fundamentals to a cohort of interns, covering both theoretical foundations and practical implementations.
• Mathematical Foundations: Guided students through essential mathematical concepts underpinning ML/DL algorithms, including linear algebra, calculus, probability theory, and statistical methods.
• Hands-on Programming: Facilitated intensive PyTorch workshops, enabling participants to implement neural networks from scratch and deploy working AI solutions.
• Knowledge Transfer: Bridged theoretical knowledge with practical applications, empowering the next generation of AI practitioners within Pakistan's developing aerospace technology sector.