Eager to work on data, pipelines, and developing end-to-end
systems.
Final-year Computer Science student at FAST-NUCES, Lahore. I investigate empirical model
robustness, design resilient data pipelines, and deploy production software that performs when the domain
shifts.
0Work Experiences across AI, Research & Software Development
0Credentials from Google, IBM, Microsoft
About
It started with a recommendation, not an equation. YouTube seemed to know
what I wanted to watch next, and I wanted to know how.
That question pulled me into computer science, then into machine learning. Today I study
why models fail when the data shifts, across scripts, domains, compression and markets, and I build the
systems that have to survive it.
I work wherever the problem needs me: researcher, data scientist, AI/ML engineer,
forward-deployed engineer, or full-stack builder. I shape data, evaluate models, design APIs, ship
interfaces, and deploy systems end to end.
“Make the complex feel inevitable: a model that earns trust, a product that reaches people, a
system that keeps working.”
Now
Research AssistantAI in Healthcare Management, FAST School of Management
Researching
Compressed CNNs vs. adversarial patchesPaper in preparation for TMLR
Building
Sachai CheckMy final year project on Urdu fake news and deepfake audio
Open to
Research labs & AI rolesSoftware, AI/ML, data and forward-deployed work
Research Interests
Models that hold up when the world shifts.
New scripts, new domains, smaller models, volatile markets. I study where models
break under real-world shift, what shortcuts they take instead of learning the task, and how to make them
transfer reliably.
اردو
Multilingual & Low-Resource NLP
Cross-lingual transfer, tokenization pathologies and bi-script dynamics across Nastaliq and Roman Urdu.
Building representative datasets where high-resource assumptions break down.
Parameter-efficient adaptation (LoRA/QLoRA) and domain-targeted instruction tuning for specialized,
low-latency models.
PEFT / QLoRAInstruction Tuning
Domain Robustness
Measuring how performance decays out of distribution, with guarded baselines that separate real
invariance from memorization.
OOD GeneralizationGuarded Baselines
Shortcut Learning
Diagnosing spurious cues such as article length or punctuation that models exploit instead of the
actual task.
Spurious CorrelationsLength Confounds
Cross-Dataset Generalization
Zero-shot transfer across corpora built for the same task, exposing when models fail to move between
dataset creators.
Zero-Shot TransferCreator Bias
2× 4× 8×
Adversarial Robustness & Model Compression
What pruning, quantization and distillation do to a model's vulnerability to adversarial patches, and
how much certified robustness survives compression. Small models get deployed; their attack surface
should be measured, not assumed.
Representation reuse from multilingual foundation models into low-resource dialects and specialized
downstream tasks.
Foundation ModelsCross-Lingual Encoders
Neural Architectures
Learnable gating and hybrid recurrent-transformer designs that adaptively fuse heterogeneous data
streams.
Dynamic GatingMultimodal Fusion
Time Series Analysis
Regime-aware forecasting that conditions temporal models on market volatility to avoid recall
collapse.
Regime SwitchingMultimodal Signals
Research & Preprints
Publications.
Preprints and papers in preparation on domain robustness, adversarial
robustness of compressed models, benchmark shortcut learning, and regime-aware multimodal fusion.
In preparation · Springer Language Resources & Evaluation (Q1)
Domain robustness across Urdu and Roman Urdu
M. A. Haroon, M. Bashir
A 26-dataset, 188,309-example benchmark in Nastaliq and Roman Urdu, released as 7
transfer matrices over 4 task types with cluster-aware splits that stop near-duplicate leakage. Across
80 domain shifts, fine-tuned encoders lose 13.5 to 47.6 macro-F1, while few-shot LLMs win 30 of 42
Holm-corrected comparisons.
Robustness of compressed CNNs on adversarial patch attacks
Does Compression Preserve Adversarial Patch Robustness? An Empirical and
Certified Study of Pruned, Quantized, and Distilled CNNs
M. A. Haroon, A. R. Antall, A. Hadi
How pruning, quantization, distillation and smaller dense networks change a CNN's
vulnerability to adversarial patches and its PatchCleanser certified robustness, across a zoo of 13
ResNet-18 models on CIFAR-10 at 2x, 4x and 8x compression. Patches made on the full-precision model
fooled the INT8 model 55% of the time, against 6–8% for random noise.
When benchmarks mislead: Reversed transfer and shortcut learning
M. A. Haroon
A zero-shot cross-dataset transfer study across Urdu fake news, English fake news,
and sarcasm detection. Fake articles in one Urdu dataset run 3.4x longer than real ones: train one way
and macro-F1 is 0.771; reverse the split and it collapses to 0.005. Models learn length, not
deception.
RAML introduces a learnable sigmoid gate that conditions the fusion of FinBERT-scored
Reddit sentiment and BiLSTM-encoded OHLCV features on rolling 24-hour market volatility. While static
concatenation causes 6-hour recall to collapse to 0.14 under volatility spikes, the RAML gate preserves
robust predictive power.
Hands-on work across research labs, healthcare AI, production software teams, and
automated AI systems.
Sep 2026 – PresentCurrent
Research Assistant
AI in Healthcare Management
FAST School of Management · Lahore, Pakistan
Surveying 50–60 doctors, then their patients, on how they use AI at work and at home, and whether
they see it as an opportunity or a threat. Supervised by Dr. Beenish Arshad.
Leading to the analysis of data in Python dashboards to scope and development of an AI healthcare
management system for scheduling, resource allocation, OPD Inventory Management, and patient–doctor
assignment, proposed for HEC NRPU.
Jul – Sep 2026
Research Intern
Urdu NLP
FAST School of Computing · Lahore, Pakistan · Supervised by
Dr. Maryam Bashir
Led two independent Urdu NLP empirical studies, written up as research papers on cross-dataset
generalization and multilingual domain robustness.
Discovered a severe length confound in fake news benchmarks causing asymmetric transfer collapse
(macro-F1 from 0.771 to 0.005) and extended analysis across English and sarcasm detection.
Built a 26-dataset, 188,309-example Urdu benchmark in Nastaliq and Roman Urdu with cluster-aware
splits so near-duplicates never leak between train and test.
Across 80 domain shifts, fine-tuned encoders lost 13.5–47.6 macro-F1, while few-shot LLMs won 30 of
42 comparisons under Holm-corrected bootstrap tests over 3 seeds.
Jun – Jul 2026
IT Intern
Enterprise AI & .NET
QuickFuel · Lahore, Pakistan
First engineer in company history to introduce AI and ML capabilities into a traditionally
enterprise .NET codebase.
Audited the central corporate portal, identifying 9 high-impact AI automation opportunities and 6
usability bottlenecks, formally presented to executive leadership.
Engineered a semantic customer sentiment prototype and ChromaDB vector retrieval layer while
contributing production C# and .NET backend code.
Jun – Jul 2026
AI Automation Intern
LLM Agents & RAG
Optimus Automate · Remote
Architected an autonomous LLM email triage agent integrated with the Gmail API, automatically
categorizing inbound correspondence and generating grounded context drafts.
Developed a production RAG document Q&A tool that strictly verifies and grounds every response
against retrieved enterprise source documentation.
Optimized context window token usage and latency through prompt engineering and vector retrieval
caching.
Featured Work
An idea does not count until something is running.
Selected production builds, RAG pipelines, autonomous systems, and evaluation
benchmarks.
01 / Sachai Check · Final Year Project
Truth checks for Urdu news and voices
Sachai means truth. Sachai Check detects Urdu fake news
across domains and deepfake audio in two pipelines, with an LLM transliterating Roman Urdu first. An LLM
judge backs each verdict with evidence and a confidence score instead of a bare real/fake label.
PythonFastAPIReactHugging FacePostgreSQL
2pipelines: text & audio
LLM judgeevidence + confidence
Ongoingfinal year project
In development
02 / TaxChronos · 2026
Stop digging through tax documents
Pakistani tax queries mix English and Urdu in the same sentence, and manual
document search took 2 to 3 minutes per question. A RAG pipeline on Qdrant and Gemini answers in under
20 seconds, scored with RAGAS against a naive-retrieval baseline.
Based on the Federal Board of Revenue (FBR) Circular 2024-25, the
applicable slab for your income bracket is 15% with a basic exemption up to PKR 600,000.
Qdrant Vector DB · 0.94 Cosine Sim · Answered in 14s
03 / CRMind · 2026
A sales CRM that flags risky leads
Sales CRM with JWT auth, a 5-stage drag-and-drop pipeline and a 4-KPI dashboard.
Gemini rates lead risk 0–100, drafts emails and reviews pipeline health, with a 30s timeout and fallback
model.
An unattended fetch, rank, summarize, and email pipeline with a Pakistan-relevance
ranking function, live on free-tier infrastructure with per-user delivery scheduling.
An open benchmark of 26 Urdu datasets in two scripts, 188,309 examples across
four task types. Fine-tuned encoders and few-shot LLMs tested on 80 domain shifts, with cluster-aware
splits and shortcut audits built in. Fine-tuned models drop up to 47.6 macro-F1 off their home domain;
few-shot LLMs hold up far better.
Whisper Tiny, Base, and Small evaluated zero-shot on 327 code-mixed Urdu-English
recordings from 11 speakers. Switch Point Error Rate isolates mistakes precisely at language-switch
boundaries.
Outlier transfer detected 02:14Isolation Forest 0.89
Process
From first conversation to production.
One disciplined thread: frame the constraint, architect data and models in
unison, then deploy reliable systems designed to run unattended.
01
Phase 01 · Day 1
Discovery
Understand the real user bottleneck, data schemas, and non-functional constraints. Inquisitive
listening, not premature pitching.
02
Phase 02 · Week 1
Scope
Written architecture document: API contracts, data pipelines, model benchmarks, and definitive
definitions of done.
03
Phase 03 · Weeks 2 to 4
Build
Clean modular code, vector stores, models, and interfaces. Rapid iterations with continuous checkpoint
reviews.
04
Phase 04 · Week 5
Harden
Stress testing on live distribution shifts, edge-cases, and confound audits. If it can break quietly, I
catch it first.
05
Phase 05 · Week 6
Ship
Containerized, documented, and deployed with full telemetry, CI/CD pipelines, and seamless handover.
Tools & Technologies
What I build with.
Production-proven technologies across machine learning, backend pipelines, and
modern web interfaces.
Frontend
React
TypeScript
JavaScript
Tailwind CSS
HTML5
CSS3
Backend & APIs
Python
FastAPI
Node.js
Express
.NET Core
C#
AI & Machine Learning
PyTorch
Transformers
TensorFlow
scikit-learn
LangChain
Whisper
Data & Infrastructure
PostgreSQL
Qdrant
MongoDB
Docker
Git
GitHub Actions
Academic Background
Education.
Rigorous computer science fundamentals and engineering leadership at Pakistan's
premier computing university.
FAST-NUCES
National University of Computer & Emerging Sciences, Lahore
Bachelor of Science in Computer Science (BSCS)August 2023 – Expected June 2027 · Final-Year Undergraduate
Key Highlights &
Leadership:
Final Year Project (FYP): Sachai Check, Urdu fake-news detection across domains and
deepfake-audio verification, with an LLM judge that explains every verdict.
Team Lead & Scrum Master: Directing cross-functional semester sprint teams across
data pipelines and full-stack deployments.
Leadership: Deputy Head of PR, FAST Business Club; Executive Media Officer, SOFTEC
(Pakistan's largest student IT Olympiad).
React
TypeScript
JavaScript
Tailwind CSS
HTML5
CSS3
Python
FastAPI
Node.js
Express
.NET Core
C#
PyTorch
Transformers
TensorFlow
scikit-learn
LangChain
Whisper
PostgreSQL
Qdrant
MongoDB
Docker
Git
GitHub Actions
