Muhammad Arsalan Iltaf — AI Researcher & Engineer

AI Researcher · NLP & Requirements Engineering

Muhammad Arsalan Iltaf

Building intelligent systems, solving complex real-world problems, and transforming research into practical innovation. I am an AI Researcher with expertise in Machine Learning and Natural Language Processing, with research published in IJIST, a journal recognized by Pakistan's Higher Education Commission. Explore my portfolio to discover my projects, publications, technical expertise, and professional journey.

MS Thesis Publications

// F1 Score (Best)

thesis ML results

// journal published

0×

HEC-recognized IJIST

// requirements analysed

annotated dataset

// detection models

0×

ambiguity smell types

Technical Arsenal

NLP & AI Research

NLP DistilBERT Transformers Text Classification Feature Engineering TF-IDF Embeddings RLHF

ML / Classical Models

XGBoost Random Forest Scikit-learn SVM Logistic Regression Explainable AI SHAP

Deep Learning

PyTorch HuggingFace TensorFlow LangChain RAG LlamaIndex LoRA / QLoRA

Research & SE

Requirement Engineering Software Quality Analysis Research Methodology Data Annotation Academic Writing LaTeX

Languages & Tools

Python SQL Git Jupyter Docker

Data & MLOps

Data Analysis Weights&Biases MLflow Pandas NumPy Matplotlib FastAPI

Projects

PROJECT_001 — FLAGSHIP · MS THESIS · PYTHON

Anaphoric Ambiguity Detection

Hybrid NLP + ML pipeline for detecting anaphoric ambiguity in natural language software requirements. Integrates DistilBERT semantic embeddings with handcrafted linguistic features (pronouns, modal verbs, vague expressions, passive voice, structural cues) into an XGBoost classifier. Hyperparameter-tuned with Optuna, threshold-optimized for class imbalance. Evaluated on the DAMIR-PURE dataset (ICSE 2022). Outputs standard + optimized classification reports, confusion matrices, and a saved model.

DistilBERT XGBoost NLP Optuna SentenceTransformer Scikit-learn DAMIR-PURE Python

View on GitHub

0.95

Best Weighted F1

10+

Linguistic Features

3-fold

Optuna CV Tuning

PROJECT_002 — NLP RESEARCH · PYTHON

Unverifiability Detection

Automated detection of unverifiability smell in natural language software requirements. Uses weighted TF-IDF representations combined with handcrafted linguistic indicators (modal verbs, vague expressions) and a Random Forest classifier. Evaluated on real-world requirements datasets with threshold optimization to handle class imbalance.

Random Forest TF-IDF Linguistic Features Scikit-learn Python MS Thesis

View on GitHub

PROJECT_003 — NLP RESEARCH · PYTHON

Referential Ambiguity Detection

Detection of referential ambiguity in software requirements using a weighted TF-IDF + handcrafted feature pipeline with Random Forest classification. Referential ambiguity occurs when noun phrases or references lack clear, unique resolution — a common quality smell in real-world NL specifications.

Random Forest TF-IDF Requirements Engineering Scikit-learn Python MS Thesis

View on GitHub

PROJECT_004 — AI APPLICATION · PYTHON

AI Resume Screening

An AI-powered resume screening system that automatically analyzes and ranks candidate resumes against job descriptions using NLP and machine learning. Extracts key skills, qualifications, and experience from resumes, computes semantic similarity scores, and outputs a ranked candidate shortlist — reducing manual screening effort significantly.

NLP Semantic Similarity Resume Parsing Scikit-learn Python HR-Tech AI

View on GitHub

PROJECT_005 — MACHINE LEARNING · JUPYTER

Iris Classification

Classic multi-class classification project on the Iris dataset using supervised ML algorithms. Demonstrates data preprocessing, visualization, model comparison (KNN, Decision Tree, SVM), cross-validation, and performance evaluation with accuracy, precision, recall, and F1-score metrics.

Classification KNN Decision Tree SVM Scikit-learn Jupyter Notebook

View on GitHub

// View all 8 repositories on GitHub

github.com/arsalaniltaf →

MS Thesis Research

MS Computer Science (AI) — 2025 NLP Requirements Engineering Explainable AI

"Automated Detection of Ambiguities in Natural Language Software Requirements"

// Abstract

Natural language remains the predominant medium for expressing software requirements, yet it is inherently susceptible to quality issues that affect precision and clarity. Among these issues, anaphoric ambiguity, referential ambiguity, unverifiability, and subjectivity appear frequently across real-world specifications and require systematic handling to support early requirements quality assessment.

This thesis develops four dedicated detection approaches — one for each smell type — based on a combination of natural language processing techniques and machine learning models. The methodology integrates contextual embeddings, weighted TF-IDF representations, and targeted linguistic indicators such as pronoun usage, modal verbs, vague expressions, and structural cues.

Anaphoric ambiguity is detected using a hybrid pipeline combining DistilBERT embeddings, linguistic features, and optimized XGBoost classification. Referential ambiguity and unverifiability use weighted TF-IDF plus handcrafted linguistic features with Random Forest classifiers. Subjectivity is handled using enhanced rule-based detection with optional ML classification.

Experiments were conducted on DAMIR-PURE, GitHub ReqEval dataset, and Zenodo datasets, yielding weighted F1-scores ranging from 0.88 to 0.95 with improved performance through threshold optimization — producing an efficient and interpretable AI-based detection framework.

// Methods & Technologies

DistilBERT XGBoost Random Forest TF-IDF NLP Feature Engineering Requirement Engineering Explainable AI DAMIR-PURE GitHub ReqEval Zenodo Datasets Threshold Optimization

0.88–0.95

Weighted F1 Range

Smell Types Detected

Datasets Evaluated

Research Publications

★ HEC-Recognized Journal Peer-Reviewed Open Access ISSN 2618-1630

A Comparative Study of Automated Approaches for Detecting Subjectivity and Unverifiability in Natural Language Software Requirements

M. A. Iltaf and A. Nadeem

International Journal of Innovations in Science and Technology (IJIST)

Vol. 8, No. 3, Pages 75–86 · 2026

// Abstract

Subjectivity and unverifiability in natural language software requirements undermine clarity and testability, making reliable automated detection valuable for early requirements quality assurance. This study presents a controlled comparative evaluation of four detection approaches: rule-based lexicons, classical machine learning models trained on TF-IDF features, a fine-tuned DistilBERT transformer, and a feature-level hybrid model that combines rule-based linguistic indicators with DistilBERT contextual embeddings.

All approaches were evaluated on a single manually annotated dataset of 985 requirements — comprising 259 subjective and 726 objective items, and 165 unverifiable and 820 verifiable items — using identical preprocessing, an 80:20 stratified train-test split, and consistent evaluation metrics to ensure a fair comparison.

The rule-based approach achieved a precision of 0.67 but near-zero recall on unverifiability, while classical machine learning models reached F1 scores between 0.47 and 0.58. DistilBERT raised weighted F1 to 0.74 for subjectivity and 0.86 for unverifiability, and the hybrid model achieved the strongest overall result, reaching a subjectivity F1 of 0.79. These findings indicate that combining linguistic cues with contextual embeddings yields the most robust performance under class imbalance.

0.79

Hybrid F1 (Subjectivity)

0.86

Hybrid F1 (Unverifiability)

985

Annotated Requirements

Models Compared

Certifications

🧠

Deep Learning using TensorFlow

IBM

COMPLETED

☁️

Machine Learning, Al & Data Science

Ehunar Punjab

2026

CERTIFIED

About

bash — arsalan_profile.sh

~$cat about.txt

AI researcher who believes that clarity in language is a prerequisite to

reliable software. I build systems that detect, explain, and fix ambiguity

in natural language — so machines and humans understand requirements the same way.

~$python -c "import arsalan; arsalan.print_profile()"

{"degree": "Master of Science (Major in Computer Science) 2025", "bs": "Software Engineering 2023",

"research": ["NLP", "Requirements Engineering", "Explainable AI"],

"published": True, "journal": "IJIST (HEC-Recognized)",

"open_to_work": True, "goal": "AI Research / Engineering role"}

~$git log --oneline -3

f2a91bc feat: threshold optimization boosts F1 to 0.95

a3c72de paper: published in IJIST Vol.8 No.3 April 2026

b1d44fa thesis: MS (Computer Science) completed — ambiguity detection

~$▌

2023 — 2025

MS Computer Science

COMPLETED 2025

Master of Science — Major in Computer Science

Specialized in NLP, machine learning, and software requirements engineering. Thesis on automated ambiguity detection achieved weighted F1 of 0.88–0.95 across four smell types. Published research in IJIST, an HEC-recognized journal.

2018 — 2023

BS Software Engineering

COMPLETED 2023

BS Software Engineering

Comprehensive foundation in software engineering principles, algorithms, databases, and systems design. Built strong programming fundamentals in Python, Java, and SQL. Final year project focused on AI-integrated software tools.

2026

IJIST Journal

PUBLISHED

Research Publication — HEC-Recognized Journal

Published comparative study on automated detection of subjectivity and unverifiability in NL software requirements. International Journal of Innovations in Science and Technology (IJIST), Vol. 8, No. 3, 2026. Co-authored with Professor Dr. Aamer Nadeem.

Contact

Open to AI research roles, NLP engineering positions, and academic collaborations.
Response time: < 24 hours · I read every message.

hello@arsalaniltaf.com

⌥

GitHub

github.com/arsalaniltaf

linkedin.com/in/muhammad-arsalan-iltaf

🎓

Journal

IJIST Vol.8 No.3 — 2026