Hafiz Muhammad Noor
BS Computer Science · COMSATS University Islamabad
Results-driven Machine Learning Engineer and Data Scientist with a strong foundation in predictive modeling and applied deep learning. Expertise spans building robust computer vision and risk forecasting models to architecting fast, late-interaction search engines using Rust and Vision-Language Models. Adept at full-stack ML engineering, from fine-tuning architectures and calibrating probabilities to optimizing hardware-level SIMD compute kernels for massive inference speedups.
GitHub Contribution Heatmap
Featured Projects & Systems
Production architectures, Rust SIMD search kernels, neural recommendation systems, and Kaggle research.
Core Work
Visual document retrieval and late-interaction (MaxSim) engine combining Vision-Language embeddings with an in-register AVX2/FMA Rust SIMD kernel scoring >90k pages/s.
Built on Coursera scraper data. Implemented collaborative filtering, KNN, baseline models, and PyTorch deep embeddings.
Frequently Asked Questions
BBQ-RAG implements a fused AVX2/FMA SIMD MaxSim compute kernel in Rust (PyO3) with 4-way query token unrolling and in-register maximum tracking. By computing late-interaction token scoring directly within 256-bit SIMD registers without allocating intermediate score matrices, it delivers 4×–8× speedups over standard PyTorch baselines, reaching 90,737 pages/second on multi-core CPUs.
I developed a multi-horizon discrete-time survival modeling framework with LightGBM to forecast right-censored wildfire escalation risks at 12h, 24h, 48h, and 72h windows. By applying isotonic regression to enforce temporal monotonicity across prediction horizons and validating via 5-fold Stratified K-Fold, the model scored 0.96 (weighted C-index + Brier score), ranking 131st out of 1,754 international teams (top 8%).
I designed a PyTorch fine-tuning pipeline utilizing ConvNeXt-Tiny and EfficientNet-B0 architectures to predict vegetative biomass metrics from imagery. Using Stratified K-Fold cross-validation with early stopping and model checkpointing to prevent overfitting, the ConvNeXt pipeline achieved a weighted R² of 0.53 on the private competition leaderboard.
Peeler features a config-driven dispatcher architecture in Python (asyncio) that isolates execution pipelines across websites, XML feeds, and PDFs. For image-based and scanned PDFs where traditional parsers fail, it integrates UTRNet (a deep learning-based Urdu OCR model via Hugging Face), paired with Rich CLI logging and graceful cancellation handling.
My core technical stack includes Python, Rust, SQL, and Bash, alongside PyTorch, Scikit-learn, LightGBM, XGBoost, and CatBoost. I specialize in computer vision, survival analysis, probabilistic forecasting, and high-throughput SIMD systems. I hold verified credentials in the IBM AI Engineering Professional Certificate and IBM Machine Learning Specialization, and am pursuing a BS in Computer Science at COMSATS University Islamabad.