Contact
Aspiring AI/ML researcher

Kashif Mahmood

AI/ML researcher · data scientist · software engineer

Software engineering graduate building a research career in computer vision, self-supervised learning, and explainable AI, with three manuscripts in revision or under review and an original, published dataset. Alongside this research, I've built production ML systems in industry, including an automated air-quality forecasting pipeline and a churn-prediction model that measurably improved customer retention. I'm currently a research assistant at the University of Sargodha, working toward a Master's in AI, machine learning, or data science.

Explore research View projects Resume
Computer vision Self-supervised learning Medical imaging
Agricultural AI MLOps
Kashif Mahmood
3
manuscripts authored or co-authored
1
public dataset (8,944 images)
1
open-source research tool

About

Who I am, what I'm working on, and the skills behind it.

Who am I

I'm a software engineering graduate with a deep interest in AI, machine learning, and intelligent systems. Over the past two years I've worked across agricultural and medical computer vision, turning research questions into datasets, models, and reproducible experiments.

Current focus

Right now, I'm focused on explainable AI (XAI): learning how to interpret and build trust in machine learning models, work that matters as much for deploying AI responsibly as for advancing the research itself. Furthermore, I recently published PakGroundnut-4V, a peanut classification dataset I designed and collected from the ground up, on Zenodo.

What drives me

I care about AI that's actually safe and reliable to deploy, not just accurate on a benchmark, whether that's a research model or a system running in production. Moving from which model scores highest to which model can we trust under real conditions is the thread running through both my research work.

Where I'm headed

My goal is a Master's in AI, machine learning, or data science, followed by a career that combines research depth with the ability to build and ship reliable, practical AI systems, whether in a research lab or in industry.

Languages & core

PythonC++SQLPHP

ML & deep learning

PyTorchScikit-learnTransfer learningCNNsVision transformersSimCLRMasked autoencoders

Data science

PandasNumPyFeature engineeringStatistical analysisCross-validation

Deployment & tooling

FlaskStreamlitGitHub ActionsHopsworksMySQL

Research

Publications, datasets, and tools built as part of my research work.

A Deployment Risk Score Framework for Self-Supervised Chest X-Ray Classification: Calibrated Multi-Objective Evaluation Under Annotation Scarcity and Distribution Shift In revision
Kashif Mahmood, Romana Aziz, Muhammad Ramzan, Mahwish Ilyas, Ala Saleh Alluhaidan
Scientific Reports · 2026 · Nature Portfolio

This study asks which pretraining approach is actually safest to deploy for chest X-ray diagnosis when labeled data is scarce and the data shifts, rather than just which one scores highest.

  • Compared seven models (SimCLR + ResNet-50 and MAE + ViT-S/16 self-supervised pretraining, and five supervised baselines including an architecture-matched ViT-S/16 trained from scratch) on a 17,988-image, six-class chest X-ray benchmark at 10%, 20%, and 100% annotation budgets, over three seeds.
  • SimCLR + ResNet-50 reached macro F1 = 0.952 and retained 88.8% of full-label performance with 10% of the labels, a result confirmed when pretraining was restricted to the training split. The from-scratch ViT-S/16 control showed that MAE's corruption robustness and low overconfidence come mainly from the ViT architecture rather than from MAE pretraining.
  • Proposed a Deployment Risk Score (DRS) that turns accuracy and robustness to eight synthetic perturbations into scenario-specific model choices, reported with calibration, cross-dataset transfer to ChestMNIST, weight-sensitivity analysis, and dataset leakage audits.
Lead authorIn Collaboration with the Authors of Princess Nourah bint Abdulrahman University
Deep Learning-Based Disease Detection in Peanut Cultivars Utilizing Transfer Learning Under review
Kashif Mahmood, Qaisar Abbas, Ume Habiba, Mahwish Ilyas, Turki Alghamdi, Hani Almoamari, Muhammad Ramzan
Scientific Reports · 2026 · Nature Portfolio

After comparing EfficientNet-B0, EfficientNet-B4 and ConvNeXt-Base with ResNet-50, I improved a ResNet-50 transfer-learning model with augmentation, a OneCycleLR schedule, and a weighted loss for class imbalance.

  • Achieved 97.28% accuracy on a 1,720-image benchmark of healthy and diseased groundnut leaf images, surpassing the previous best published result of 96.51% on the same dataset.
  • Addressed class imbalance and small-data training on a five-class leaf dataset, with the full pipeline and results reported in a public preprint.
Lead authorIn Collaboration with the Authors of Islamic University of Madinah
A Systematic Review of Visual Computing Technologies for Intelligent Peanut Agriculture Under review
Ume Habiba, Nabeel Abbas, Kashif Mahmood, Mahwish Ilyas, Muhammad Summair Raza, Muhammad Ramzan
USJICT · 2026 · University of Sindh, Jamshoro

A review of deep learning, machine learning, and hybrid approaches across four core tasks in peanut agriculture: variety classification, maturity assessment, disease detection, and pest recognition.

  • Compared benchmark datasets and evaluation metrics across the literature, identifying dataset scale, model complexity, and field adaptability as the primary open challenges.
  • Proposed future research directions, including lightweight visual models, multi-task integrated systems, and soil-penetrating sensing technologies.
Corresponding authorIn Collaboration with the Authors of University of Sargodha
PakGroundnut-4V: A Multi-Condition Image Dataset for Pakistani Groundnut (Arachis hypogaea L.) Variety Classification Published

A large-scale, multi-condition image and phenotypic dataset built to support robust computer vision research in agricultural variety classification. Designed and collected independently, with images gathered across two sites in Punjab, Pakistan: Chakwal and Sargodha.

Varieties Attock-2019 (2,240 images), Fakhr-e-Chakwal (2,288), Parachinaar (2,192), Fakhr-e-Attock (2,224); verified through peer-reviewed literature and direct consultation with the Barani Agricultural Research Institute (BARI), Chakwal
Scale 8,944 images (320×320 px), fixed 90° overhead angle, across 5 backgrounds, 4 lighting conditions, and 5 quantity levels (1 to 5 peanuts per frame)
Negative class 60 background-only images, kept separate from the classification folders to support out-of-distribution and future object-detection research
Morphometric samples 80 pod photographs and measurements (20 per variety), taken with a digital Vernier caliper (0.01mm precision) and digital scale (0.1g precision)
Splits Standard stratified split (6,259 / 1,340 / 1,345 train/val/test), plus dedicated cross-background and cross-lighting splits for domain-generalization research
Baseline Five models over five seeds. ViT-B/16 reaches a mean macro-F1 of 0.977 (best seed 0.982) on the standard split, but drops by 22.3% on unseen backgrounds and 17.4% on unseen lighting
Documentation README, LICENSE, CITATION.cff, and DATA_CARD.md, with full per-image metadata
License CC BY 4.0
Access Zenodo (DOI: 10.5281/zenodo.22014996), Kaggle, GitHub
Authors Kashif Mahmood, Dr. Muhammad Ramzan, Khizar Hayat (farm owner who supplied physical specimens)
Zenodo Kaggle GitHub
Image Dataset Cropper Public tool

A full-stack local web application built to solve the bottleneck behind preparing PakGroundnut-4V: batch-cropping composite dataset photographs into individual, ML-ready images.

Stack Python 3.11, FastAPI, Pillow, React 18, Vite
Core feature Custom HTML5 canvas multi-region crop editor with drag, resize, and one-click batch-apply across a folder
Engineering Chunked uploads, streaming ZIP export, disk-cached thumbnails — processes multi-gigabyte datasets with no size limit
View repository

Experience

Research and applied data science roles.

Sep 2024 – present
Research Assistant
University of Sargodha, under Dr. Muhammad Ramzan
  • Served as lead or co-author on three research manuscripts spanning agricultural and medical imaging AI, including a deployment-focused evaluation of self-supervised learning for chest X-ray classification and a transfer-learning approach to peanut disease detection, both at Scientific Reports.
  • Independently conceived, built, and documented a public image dataset of 8,944 peanut images, including custom-built data-preparation software.
  • Conducted literature reviews, experimental design, and model development across deep learning and self-supervised learning projects, in collaboration with international research partners at the Islamic University of Madinah and Princess Nourah bint Abdulrahman University.
View letter
Oct 2025 – Nov 2025
Data Science Intern
10Pearls, Remote (part-time, concurrent with research assistantship)
  • Built a fully automated, end-to-end 72-hour Air Quality Index (AQI) forecasting system for Karachi's NED University region,from raw API ingestion through deployment.
  • Engineered the machine learning pipeline: feature engineering, model evaluation across four algorithms via GridSearchCV,and multi-step forecasting via a MultiOutputRegressor, with all features and models versioned in Hopsworks.
  • Automated the entire pipeline with dual GitHub Actions workflows for hourly data ingestion and daily model retraining,requiring no manual intervention.
View letter
Jun 2024 – Aug 2024
Data analytics intern
BrandPur, Sargodha
  • Analyzed customer data to understand retention and churn behavior, using Python (pandas, matplotlib, seaborn) for analysis and Power BI for dashboarding.
  • Built a churn-prediction model estimating the probability of customer attrition, which informed targeted retention packages for at-risk customers.
  • Contributed to an approximate 10% increase in customer retention through this data-driven retention strategy.
View letter

Education

Bachelor of Science in Software Engineering

University of Sargodha
  2021 – 2025
CGPA: 3.90 / 4.00

I built a strong foundation in Programming Fundamentals, Object-Oriented Programming, Data Structures and Algorithms, Statistics, Linear Algebra, and Database Management Systems, alongside hands-on experience with essential tools. These fundamentals gave me the confidence to work with Machine Learning and Deep Learning models. In my fourth semester, I independently began exploring Data Science, the starting point of my path into AI.

Intermediate in Computer Science (ICS)

Superior College, Chiniot
  2019 – 2021
Marks: 1013 / 1100 (92.09%)

I completed my Intermediate studies in Computer Science with strong academic performance, building a foundation in programming, computational theory, and problem-solving. This period sparked a genuine interest in technology that pushed me to explore concepts beyond the syllabus on my own, refining my skills and deepening my understanding of the technical landscape. That early curiosity gave me clear direction toward a career built around technology and innovation.

Portfolio

A mix of research systems, applied data science, and full-stack builds.

Achievements

Academic honors, leadership, seminars, sports, and certifications — filter by category.

Also available for data science consulting

Outside of research, I take on freelance and contract work in data preprocessing, ML model development, dashboards, and CI/CD-automated pipelines.

Data cleaning & EDA ML model development Dashboards CI/CD automation
Get in touch

Get in touch

Open to research collaborations, graduate supervision conversations, and select freelance work.

Email
kashifmahmood.scholar@gmail.com
Phone
+92 337 1733897
Location
Chiniot, Pakistan