Architecting AI. Production Engines.

Engineering high-throughput computer vision engines, asynchronous OCR extraction pipelines, and PyTorch deep neural recommendation models for enterprise production.

GitHub Contribution Activity@cmd-HMN ↗

GitHub Contribution Heatmap

SepOctNovDecJanFebMarAprMayJunJulAugSepMonWedFri

Education & Skillset

Computer Science scholar at COMSATS University Islamabad specializing in end-to-end Machine Learning pipelines, model optimization, and high-throughput data extraction engines.

PROFILE ABSTRACT

Hafiz Muhammad Noor

Dedicated to solving complex engineering challenges through mathematical modeling and machine learning. Deeply engaged in the Python data ecosystem—building predictive architectures, computer vision models, asynchronous scraping systems, and chess logic engines from scratch.

ML ResearchComputer VisionUrdu OCRScikit-Learn
ACADEMIC FOUNDATION

COMSATS University Islamabad

Bachelor of Science in Computer Science
Undergraduate Program

Technical Mastery & Frameworks

Core Languages

Python
Primary
Rust
Systems
C++C++
Performance
Bash
Automation
SQL
Database

ML Frameworks & Libraries

PyTorch
Deep Learning
Scikit-learn
Predictive ML
TensorFlow
Neural Nets
Pandas
Data Analysis
FastAPI
Model APIs

Infrastructure & MLOps

Docker
Containers
AWS
Cloud Compute
GITGit
VCS
LangChain
LLM Agents

IBM Professional Certifications

IBM | Coursera
Jun 2025

AI Engineering Professional Certificate

Mastered deep learning models, generative AI architectures, neural network training with PyTorch & TensorFlow, NLP, and model deployment strategies.

View Verified Certificate
IBM | Coursera
Mar 2025

Machine Learning Specialization

Comprehensive training in supervised & unsupervised learning, gradient boosted decision trees (XGBoost), model validation metrics, and Scikit-learn pipelines.

View Verified Certificate

Projects & ML Systems

A collection of open-source models, data pipelines, Kaggle research, and algorithmic experiments.

Featured Work

Kaggle Competitions & ML Notebooks

Frequently Asked Questions

How does BBQ-RAG achieve over 90,000 document pages scored per second on CPU?

BBQ-RAG implements a custom Rust AVX2/FMA SIMD compute kernel (`maxsimd`) with 4-way query token unrolling and in-register maximum tracking. By computing late-interaction MaxSim dot-products directly inside 256-bit YMM registers without intermediate matrix allocations, it cuts memory traffic by 27x and scales linearly with Rayon adaptive work pools.

What is your primary machine learning framework workflow?

I leverage PyTorch for deep neural network prototyping, custom loss formulation, and GPU model optimization. For structured tabular predictive challenges, I rely on Scikit-learn, XGBoost, and LightGBM cross-validation pipelines.

How does the Peeler Urdu OCR extraction pipeline handle high throughput?

Peeler uses Playwright's asynchronous event loop to execute multi-threaded browser sessions in parallel. Extracted image regions undergo adaptive threshold binarization before passing through EasyOCR models.

How do you approach data cleaning and feature engineering for noisy datasets?

I apply robust outlier detection via IQR/Isolation Forests, categorical target encoding, imputation strategies based on feature distributions, and synthetic balance (SMOTE) where class imbalance degrades precision-recall metrics.

How are trained models exposed to web clients and external microservices?

Models are serialized via TorchScript or ONNX runtime and wrapped inside asynchronous FastAPI endpoints. Docker containers are deployed to cloud infrastructure with automated health probes and latency monitoring.