Abdullah
Final year AI student at FAST NUCES with hands on experience designing and deploying intelligent AI systems. Specializes in LLM powered applications, agentic AI frameworks, and RAG pipelines. Co authored an IEEE accepted research paper achieving a 9.1 percent improvement in multi hop question answering and 52 percent reduction in hallucinations using a custom multi agent framework. Proven ability to translate complex business requirements into robust, automated AI solutions using LangChain, LangGraph, FastAPI, and modern vector databases.
Experience
Machine Learning Intern
Nebunex AI
Islamabad
Designed and deployed LLM powered pipelines for clinical text analysis using transformer architectures, prompt engineering, and embedding techniques to extract structured medical insights from unstructured data. Built a multimodal AI system integrating ECG imagery with clinical reports via the OpenAI CLIP framework, leveraging vision language alignment and contrastive learning to enhance diagnostic prediction accuracy. Engineered scalable data pipelines for MIMIC III and MIMIC IV healthcare datasets, processing over 50,000 patient records using Python, Pandas, and NumPy with parallel processing to accelerate large scale model training. Developed and optimized ML models with Scikit Learn through cross validation, hyperparameter tuning, and ensemble methods, achieving measurable improvements in prediction accuracy benchmarks.
Education
National University of Computer and Emerging Sciences (FAST)
Bachelor of Science
Relevant Coursework: Machine Learning, Deep Learning, Natural Language Processing, Computer Vision, Generative AI, Agentic AI, Digital Image Processing, Database Systems, Computer Networks
Skills
Research Publication
Agentic Self-RAG: Multi-Agent Reasoning for Self-Correcting Retrieval Augmented Generation
IEEE ICIT 2026(Accepted)
Co authored a multi agent orchestration framework that addresses distinct RAG failure modes through three role specialized agents: a Query Analyzer for question decomposition and ambiguity detection, a Retrieval Critic for evidence sufficiency and contradiction detection, and an Answer Verifier for faithfulness scoring and hallucination flagging, coordinated via LangGraph with explicit confidence thresholds. Achieved 54.3% exact match on the HotpotQA validation set (7,405 samples), a 9.1 percentage point improvement over Vanilla RAG, and 68.5% accuracy on StrategyQA, a 9.3 percentage point gain; hallucination rate dropped by 52% and faithfulness score reached 0.82 BERTScore. Confirmed statistically significant improvements through paired t tests and Wilcoxon signed rank tests (p< 0.001), with Cohen’s d = 0.251 on HotpotQA and Cohen’s d = 0.287 on StrategyQA; ablation study showed the Answer Verifier alone contributed a 4.1 percentage point gain in exact match.
Projects
AI Powered Metro Bus Tracker— Full Stack Agentic Platform
Built an AI-powered metro bus tracking platform for Islamabad using agent-based simulation; each bus acts as an autonomous agent with real-time traffic-aware Estimated Time Arrival prediction (time-of-day factors, stochastic traffic events) and live congestion visualization via color-coded route segments. Developed a React Native mobile app and React.js admin dashboard with WebSocket-based live fleet monitoring, OpenStreetMap, and full CRUD management over a FastAPI + MySQLite backend, and a chatbot supporting for commuter queries on routes, fares, and delays. Designed an Intelligent Journey Planner supporting 0–3 transfers via progressive DFS with backtracking, geodesic nearest-stop discovery, and a weighted scoring model (time 40%, transfers 20%, walking 15%) returning top-5 options with a RECOMMENDED tag. Tools:React Native, React.js, FastAPI, LangChain, WebSocket, OpenStreetMap, Docker, GitHub Actions
Multi Tenant Agentic RAG System
Developed a production grade multi tenant RAG system using LangChain and ChromaDB with a multi layered security framework, including prompt injection detection, ACL based tenant isolation, and PII masking. Integrated Groq LLM with mandatory citation enforcement and conversational memory to deliver context aware, verifiable responses across fully isolated tenant environments. Tools:LangChain, ChromaDB, Groq API, Python, FastAPI
Text to Image Generation Microservice
Built a containerized gRPC microservice for AI driven image generation using HuggingFace diffusion models, with concurrent request handling and a Gradio frontend; evaluated open source models across latency, quality, and cost trade offs to select the optimal production model. Tools:HuggingFace, PyTorch, gRPC, Docker, Gradio, Postman
Image Classification System— CNNs and Vision Transformers
Developed a web based image classification system using Convolutional Neural Networks and Vision Transformers to deliver AI powered predictions with confidence scores, backed by a Django REST backend and an interactive HTML, CSS, and JavaScript frontend. Tools:PyTorch, TensorFlow, OpenCV, Django
Certifications
How Transformer LLMs Work
DeepLearning.AI
Gained practical understanding of transformer architecture, attention mechanism, tokenization, and training of large language models. Covered concepts such as self attention, positional encoding, and real world applications of LLMs in natural language processing.