The University of Florida — PhD in Computer Science
Aug 2025 – Present • Advisor: Prof. Guanpeng Li • Focus: dependable HPC, fault tolerance, GNN, DBN , SDC modeling, LLVM, Fault Injection.
Current DATA SCIENCE • FAULT TOLERANCE • ML
Data Scientist. Fault Tolerance
Research-driven ML engineer & PhD student: dependable systems for high-performance computing, fault tolerance, LLVM-based fault injection, and silent data corruption modeling. I make complex pipelines feel simple.
Education
The University of Iowa — PhD in Computer Science
Aug 2024 – 2025 • Advisor: Prof. Guanpeng Li • Focus: dependable HPC, fault tolerance, LLVM-based fault injection & SDC modeling.
Transfered CGPA 4.0LUMS — BSc Computer Science
2019 – 2023 • CGPA 3.78 • Dean’s Honor List (2019–2022). Core: ML, DL, AI, Speech, Data Mining, HCI, Security, SE, Automata, NCC.
Honors CGPA 3.78Skills & Abilities
Machine Learning & Data Science
PyTorch85%
TensorFlow80%
Scikit-learn75%
Pandas / NumPy85%
Development
JavaScript (ES6+)80%
React / React Native85%
HTML5 / CSS375%
Programming Languages
Python90%
C++80%
SQL75%
Systems & Tools
LLVM75%
Linux80%
Git / Docker75%
Experience
Graduate Research Assistant — University of Florida
Research on dependable ML/HPC; fault tolerance and error resilience for HPC workloads.
Research Assistant — Dependable Systems Lab, University of Iowa
LLVM/runtime tools for instruction-level resilience analysis; fault detection in HPC workloads.
Research Assistant — Dependable Systems Lab, University of Iowa
Resilience tooling and analyses; supported experiments for dependable systems research.
Teaching Assistant — Artificial Intelligence, University of Iowa
Support 30+ students with office hours, grading, and learning outcomes for AI coursework.
Research Assistant — DebloatBench (SRI • UArizona • LUMS)
A research project in collaboration with Dr. Sazzadur Rehman (Assistant Professor at the University of Arizona), Dr. Ashish Gehani (Principal Computer Scientist at SRI) and Dr Fareed Zaffar (Assistant Professor at Lahore University of Management Sciences). The project provides a unified framework to evaluate a diverse set of container debloaters that can handle the diversity of design and execution environments for container debloating.
Associate Consultant — AI, Systems Limited
ML & LLM Solutions for Global Organizations: - Collaborated with Microsoft to develop LLM solutions using Azure Cloud services. - Implemented solutions with Azure Cognitive Services, Omni Channel, RAG architecture, Databricks, and Data Robot. - Fine-tuned models using PEFT and LoRA techniques for optimal performance. - Applied multiclass classification and clustering for tailored ML solutions, integrating latest research advancements. LLM Training Sessions for Professionals: - Conducted comprehensive training on LLMs, covering Transformer architecture and RAG. - Designed and facilitated quizzes, labs, and final projects at Systems Limited. - Delivered practical, industry-focused courses emphasizing NLP basics and fine-tuning techniques.
Research Assistant — ISPL, LUMS
Fraud Detection – The project drew inspiration from existing works on credit card fraud detection and focused on using Machine learning techniques to reduce security risks in financial transactions by leveraging large datasets provided by Clariba SEIDOR, a consultancy firm • Focused on refining credit card transaction security through techniques such as ensemble learning, data preprocessing, and the application of SMOTE to address class imbalance challenges. • Executed comprehensive literature reviews, conducted exploratory data analysis, and applied and compared classification and clustering techniques for fraud detection. • Integrated Snowflake and Dataiku tools to harness scalable computational power for machine learning model development and assessment. • Addressed practical challenges in credit card fraud detection, including class imbalance and memory constraints, showcasing adaptability and solution-oriented thinking.
Data Analysis Intern — VentureDive
• GeoSpatial Data Processing o Applied Selenium for efficient web scraping to collect National and Provincial Constituencies data in Pakistan. o Processed and formatted the acquired data to prepare it for subsequent analysis and visualization. • Interactive Webpage Development o Developed an interactive webpage using Python, Flask, and HTML. o Integrated the GeoJson file generated from the collected data, enabling users to explore and interact with the geographical information easily. • Utilized online plotting tools to enhance the visual representation of geographical data on the webpage. • Implemented user-friendly features, allowing users to zoom, pan, and retrieve detailed information about specific constituencies directly from the webpage..
Branch Banking Intern — Habib Bank Limited
ML for employee-effectiveness metrics; campaign ops & performance analytics.
Peer Advisor — LUMS
Mentored ~30 students on course/major planning and academic progress.
Teaching Assistant — Computational Problem Solving, LUMS
Led tutorials for 100+ students; authored quizzes/exams; grading & office hours.
Data Science Intern — JS Bank
• Executed daily SQL queries to optimize the ETL process for efficient data extraction, transformation, and loading. • Leveraged existing customer data through SQL queries, contributing to the development of predictive analysis models. • Translated predictive insights into actionable strategies, enhancing informed decision-making across diverse organizational departments.
Projects
Enterprise RAG & PEFT on Azure
Production-ready retrieval-augmented generation with Cognitive Services, vector search, and LoRA adapters for task-specific tuning.
Stack: Azure OpenAI, Azure Search/Blob, LangChain/Prompt-flow, PEFT/LoRA, Docker.
- Chunking + hybrid retrieval (sparse+dense); guardrails for PII and prompt injection.
- Tracing & eval harness (faithfulness, context recall, answer quality) with offline test sets.
- Keywords: MLOps, observability, vector stores, red-teaming, governance.
Streaming Language Classifier
Low-latency multi-lingual classifier; distilled models with real-time constraints for edge serve.
Stack: PyTorch, TorchScript, scikit-learn, FastAPI.
- Tokenizer experiments (BPE vs WordPiece); class-imbalance handling; label smoothing.
- Exported to TorchScript; batched inference; throughput vs accuracy tradeoffs.
- Keywords: distillation, quantization, latency SLOs, A/B testing.
Knowledge Distillation Study
CIFAR-10 compression with multi-student KD; systematic ablations over T, depth, and data aug.
Stack: PyTorch, Albumentations, WandB/Matplotlib.
- Teacher-student ensembles; temperature & loss-weight sweeps; early-stopping & checkpointing.
- Keywords: model compression, calibration, robustness, learning curves.
YAPL — Yet Another Programming Language (Interpreter)
From grammar to AST to bytecode-like executor; showcases end-to-end compiler pipeline skills.
Stack: Python, PLY (lex/yacc), unittest.
- Lexer/parser, AST nodes, visitors, constant folding & dead-code elimination.
- Keywords: grammars, CFG, tokens, semantic analysis, IR.
Celebrity Identification (Classic CV)
End-to-end CV pipeline: detection → feature engineering → classification with a lightweight web UI.
Stack: OpenCV (Haar), scikit-learn, Flask, Wavelets.
- GridSearchCV for model/param selection; simple deployment artifact for demo.
- Keywords: classical ML, feature extraction, evaluation, microservice.
Knowledge-Base QA Bot
Retrieval-augmented QA with ChromaDB; adds conversation state & lightweight guardrails.
Stack: Python, ChromaDB, LangChain, FastAPI/Flask.
- Chunking & embedding; answer-faithfulness evaluation; simple UI for enterprise demos.
- Keywords: vector search, reranking, grounding, guardrails.
Interactive GeoSpatial Web
Scraped civic data → GeoJSON → interactive maps (zoom/pan/popovers) with a small Flask backend.
Stack: Selenium, Python, GeoJSON, Flask, Leaflet/HTML.
- Automated ETL & sanity checks; searchable map with constituency details.
- Keywords: data engineering, scraping, spatial joins, visualization.
In-Browser MNIST Classifier
Digit prediction fully client-side via ONNX Runtime Web (WebGL → WASM fallback); ~25–40 ms inference on desktop.
Stack: ONNX Runtime Web, Canvas API, vanilla JS.
- Path-robust model loading for GitHub Pages; graceful degradation and clear UX errors.
- Keywords: model export, client-side inference, WebGL, WASM, DX.
Draw-a-Digit (0–9)
Draw below and hit Predict.
Probabilities
Copied!
