I build and study machine learning systems for places where the predictions
actually matter — clinics, classrooms and mountain valleys.
Two peer-reviewed papers, one as first author, and a decade's worth of
software shipped to people who live three hours from the nearest city.
03WHO polio campaigns surveyed in northern Chitral
Scroll
i.
A researcher who grew up far from the lab
Chitral sits at the far north of Pakistan, wrapped in the Hindu Kush. It shapes what I choose to work on.
Background
I finished my BS in Computer Science at the University of Chitral with a CGPA of 4.00/4.00 and the Gold Medal, and published twice during my final year — once as first author.
My work compares supervised learning methods on problems where a wrong answer has a cost: cardiovascular risk, and student outcomes. Both taught me the same lesson. The interesting question is rarely which model scores highest. It is whether a model is trustworthy enough to act on when someone's health or education depends on it.
Between research and engineering I have surveyed polio vaccination coverage door-to-door for the WHO, built billing software a local pharmacy uses daily, and made a place online for poets writing in Khowar — my mother tongue.
Research focus
Human-centered AI. Computer vision, deep learning, and the fairness and reliability of models once they leave the notebook.
Now
Web Infrastructure Engineer at Cyber Castellum. Incoming Generative AI intern at the KP IT Board.
Publications
2 journal papers, 2025
Both in Spectrum of Engineering Sciences. First author on the cardiovascular study.
Languages
Khowar Native
Urdu Fluent
English Fluent
Pashto Basic
ii.
Published research
Two studies on supervised learning where the outcome affects a real person. Open either one for the method, the numbers and what I would do differently.
P1
A comparative study of machine learning algorithms for cardiovascular risk prediction: SVM, Gradient Boosting and Rotation Forest
Spectrum of Engineering Sciences, 3(8), 481–491 · August 2025 · First author
Cardiovascular disease causes roughly 17.9 million deaths a year, and in a district like mine
the diagnostic bottleneck is not treatment — it is getting to a specialist at all. I wanted to know
which classifier families hold up on routine clinical variables, and what the failure modes cost.
I built the pipeline in Python with scikit-learn: outlier and missing-value handling, feature/label
separation, a train-test split, then Support Vector Machine, Gradient Boosting and Rotation Forest
trained independently on 1,025 patient records across 14 features. Evaluation used accuracy,
precision, recall and F1, with confusion matrices read specifically for false negatives — the error
that sends a sick patient home.
1,025patient records
14clinical features
3classifiers compared
64.9%SVM accuracy
The two ensemble methods clearly outperformed SVM, which reached 64.9% accuracy with a 0.72 F1.
Rotation Forest's PCA-rotated feature subspaces and Gradient Boosting's sequential residual fitting
both handled the multivariate structure far better than a single margin-based separator.
What I would change. The ensembles scored perfectly on the held-out set, and a perfect score
is a finding about the data, not the model. The 1,025-row dataset is an expanded version of the
303-row UCI Cleveland set and contains duplicated records, so near-identical rows almost certainly
straddled the split. Re-running this properly means deduplicating before splitting and reporting
cross-validated scores. I would rather say that plainly than quote the headline number.
Predicting academic success: a machine learning approach using Decision Tables and Random Forests
Spectrum of Engineering Sciences, 3(5), 205–213 · May 2025 · Co-author
Schools can only intervene early if they can see who is slipping early. This study asked whether a
model simple enough for a school to actually read can compete with an ensemble on the same task.
We used the UCI student-performance dataset — 395 students, 33 attributes covering grades, family
background, study time, absences and social factors — and applied best-first search as a feature
selection step, which cut the space to 11 attributes including prior period grades, past failures,
mother's education and family relationships. Decision Tables and Random Forests were then trained on
that reduced set and compared on correlation coefficient, RMSE and mean absolute error.
395student records
33 → 11attributes after selection
1.92Decision Table RMSE
1.99Random Forest RMSE
The interpretable model won. The Decision Table reached a 0.92 correlation coefficient with the
lowest error on both RMSE and MAE, edging out Random Forests. For an institution with no data
science team, a model whose rules a teacher can read and challenge is worth more than a marginal
accuracy gain — which is the part of this result I keep coming back to.
Research prototypes, AI products and software that people in Chitral use. Filter by what you care about.
Research2025
Cardiovascular risk prediction
End-to-end scikit-learn pipeline benchmarking SVM, Gradient Boosting and Rotation Forest on 1,025 clinical records, with confusion-matrix analysis focused on the cost of false negatives. Published as first author.
Pythonscikit-learnPandasMatplotlib
Research2025
Academic success prediction
Feature-selection and model-comparison study on UCI student data, using best-first search to cut 33 attributes to 11 before pitting interpretable Decision Tables against Random Forests.
Decision TablesRandom ForestBest-first search
AI product2025
Negotia
A Chrome extension that scrapes live web context — listings, comparable prices, counterparty signals — and drafts grounded negotiation replies. The human always sends the final message; the model only proposes.
LLMJavaScriptChrome APIsScraping
AI product2025
Conversational CV builder
You talk, it writes. The system extracts structured résumé fields from free-form conversation and applies edits to the live document as you speak, replacing the form-filling interface entirely.
LLMReactStructured output
Product2025
Pharmacy point-of-sale, with AI
A desktop system handling billing, stock and sales reporting for pharmacy retail, plus a natural-language layer so the owner can ask what is running low instead of reading a report. Deployed and in daily use in Chitral.
PythonDesktop UIMySQLLLM
Systems2024
Blockchain certificate authentication
Degree and certificate verification built on SHA-256 hashing, so an employer can confirm a document is untampered without phoning a registrar's office that may not answer.
SHA-256BlockchainPython
Civic2024
Flood alert management system
Early warning and incident reporting for the Chitral valleys, where glacial outburst floods arrive fast and the existing alert chain is a phone call. Built after watching one of those summers.
DjangoMySQLGeo data
AI2024
Library system with chatbot
Catalogue search, issue and return handled through conversation rather than a form — aimed at students who know the book they want but not the classification system it lives under.
PythonNLPMySQL
Community2025
Khowar poetry platform
A publishing and discussion space for poets writing in Khowar, a language with roughly half a million speakers and almost no digital presence. Low-resource languages do not preserve themselves.
ReactMySQLUnicode
Systems2023
Student result portal
Secure login, result lookup and full history for a college with no budget for a student information system — built entirely on Google Sheets with an Apps Script backend so it costs nothing to run.
Apps ScriptGoogle SheetsAuth
Mobile2023
Attendance tracking app
An Android app for collecting, storing and correcting student attendance, with absence marking and editable history — replacing a paper register that got lost twice a semester.
AndroidSQLite
Web2022
WordPress builds & SEO
Freelance content sites and small-business presence work: theme customisation, plugin architecture, page performance and search visibility for clients who need to be found.
WordPressPHPSEO
iv.
Experience and education
Engineering roles, field work and a degree finished at the top of the class.
Experience
Jun 2026 — present
Web Infrastructure Engineer · Cyber Castellum
Designing, deploying and maintaining production web infrastructure, with responsibility for performance, reliability and security hardening.
2026 · incoming
Generative AI Intern · KP IT Board
Selected for a competitive government programme in applied generative AI.
Six-month national apprenticeship under the Pakistan Software Export Board, delivering client-facing web features inside a professional engineering team.
Apr 2025 — Nov 2025
Frontend Developer · Hindukush Soft Technologies
Built and shipped responsive React interfaces for client products, with API integration and code review.
Three campaigns
Polio Campaign Surveyor · World Health Organization
Field surveyor for polio immunisation drives across northern Chitral — verifying household vaccination coverage and reporting data back from remote mountain communities. The clearest lesson I have had in how messy real data collection is.
Six months
Computer Instructor · Jamiat Coaching Academy
Taught programming and computing fundamentals to secondary and intermediate students, and wrote the course material.
Education
BS Computer Science
University of Chitral · 2021–2025 CGPA 4.00 / 4.00 — Gold Medallist
HSSC — Pre-Medical & Additional Mathematics
Govt. Higher Secondary School, Drosh · 2017–2020 880/1100 and 874/1100
SSC — Science
Drosh Public School · 2015–2017 920/1100
Leadership
President, Youth Empowerment Club — University of Chitral. The club placed first among 35 youth clubs nationally under my term.
President, Cyber Force Society — ran workshops and awareness events on cybersecurity and emerging technology.
General Secretary, Student Aid Foundation for Education — fundraising and support for students facing financial barriers.
v.
Toolkit and recognition
What I work with, and what has been recognised along the way.
I am looking for an MS position where I can work on human-centered computer vision and
the reliability of deployed models. If you supervise in that area, I would like to hear
from you — and I am also available for freelance engineering.