AI/ML Engineer — Islamabad, Pakistan
I build AI systems that hold up outside the notebook.
From LLM-powered agentic workflows to computer vision pipelines running in production — I design, train, and ship AI systems end to end, backed by published research in medical imaging.
Sketchey.com
Real-time image processing, live in production
LivePixelSpark
AI photo editor running 60fps, fully offline
Open SourceClinical Attention Consistency
Vision Transformer research, published on Zenodo
Zenodo 2026Professional & research roles
Projects shipped end-to-end
Published research paper
Deepfake-detection accuracy (EfficientNet-B0 + forensics)
About
Research and production, one throughline
I’m a Computer Science undergrad at FAST NUCES (2023–2027, Dean’s List 2025), currently a Teaching Assistant and an AI Research Participant at NuSYS Lab. Most recently, as an AI/ML Intern at Xandec, I built LLM & RAG pipelines and agentic workflows with LangChain, LangGraph, and CrewAI for multi-step task automation, alongside computer vision models and the data pipelines that support them.
My research sits in medical AI: I designed a Clinical Attention Consistency (CAC) Loss that aligns Vision Transformer attention maps with radiologist annotations, improving cross-hospital generalization on 112K chest X-rays across 14 diseases — published open-access on Zenodo. Outside of that, I’ve shipped independent projects end to end: a production image-processing platform, an on-device AI photo editor, a deepfake forensics tool, and a virtual try-on pipeline.
Experience
Where the work has happened
- Jun 2026 – Aug 2026
AI/ML Intern
Xandec · Islamabad, Pakistan (Hybrid)
- Built and evaluated LLM & RAG pipelines, grounding model responses in domain-specific data
- Designed agentic workflows with LangChain, LangGraph, and CrewAI for multi-step task automation and tool-calling
- Trained and evaluated computer vision models for applied product use cases
- Built data pipelines for model training and evaluation, streamlining preprocessing and experiment tracking
- 2025 – Present
AI Research Participant
NuSYS Lab, FAST NUCES · Islamabad, Pakistan
- Designed the Clinical Attention Consistency (CAC) Loss for ViT-based medical imaging
- Authored and published an open-access research paper on Zenodo (2026)
- Jan 2026 – Present
Teaching Assistant
FAST NUCES · Islamabad, Pakistan
- Conduct weekly lab sessions for 50+ students in C++ and core programming concepts
- Hold 3 office hours weekly for one-on-one debugging and concept support
- Developed supplementary learning materials that improved class performance by 15%
- Mar 2023 – Present
Open Source Contributor
GitHub · Remote
- Contributed bug fixes and documentation improvements to Python libraries used in computer vision projects
- Resolved 5+ issues across AI/ML repositories
Work
Featured projects
Sketchey.com
Real-Time Image Processing Platform
Production-deployed full-stack image enhancement platform implementing pixel-level digital image processing as an interactive web app — pencil-sketch conversion, cartoon effects, filters, and transformations.
- Deployed on a Linux VPS behind an Nginx reverse proxy with SSL and Dockerized services
- Optimized the async processing pipeline for low-latency, real-time editing
PixelSpark
On-Device AI Photo Editor
A professional AI photo editor for Android running a real deep-learning model entirely on-device, with zero cloud dependency.
- MIRNet compressed to a ~27MB TFLite model for low-light restoration, running inference in a background isolate to keep the UI at 60fps
- Custom C++17 image-processing core bridged to Flutter via dart:ffi, plus 21+ pro editing tools with full non-destructive undo/redo
Deepfake Image Detector
Hybrid Forensics + Deep Learning
A hybrid deepfake-detection pipeline combining a deep-learning classifier with classical forensics for a more trustworthy verdict.
- EfficientNet-B0 classifier (92% accuracy) ensembled with Error Level Analysis forensics via a weighted verdict
- Deployed full-stack with real-time confidence scoring
Body Measurements & Virtual Try-On
Computer Vision Pipeline
A computer vision pipeline for body-measurement estimation and virtual garment try-on, delivered through a companion Flutter app.
- Fine-tuned SAM 3D for body segmentation and CatVTON for realistic garment fitting
- Delivered end-to-end through a Flutter mobile app for a complete try-on experience
Research & Recognition
Published, peer-facing work
AUC across 14 diseases
Clinical Attention Consistency for Cross-Hospital Robustness in Vision Transformers for Chest X-Ray Classification
- Novel CAC Loss supervises ViT-B/16 attention maps against 880 radiologist-annotated bounding boxes
- 112,120 NIH chest X-rays across 14 diseases; 71% of diseases generalized to an unseen CheXpert hospital dataset with zero retraining
- Handled extreme 637:1 class imbalance via weighted BCE, AdamW with differential learning rates, and cosine-annealing scheduling
Achievements
- Dean's List 2025 (Top 10%), FAST NUCES
- Trilliet AI Hackathon 2025 — Finalist (top teams of 50+)
- Teaching Assistant, FAST NUCES (Spring 2026 – present)
- Open Source Contributor — 3+ years, 5+ issues resolved across AI/ML repos
Writing & Community
- 5+ technical articles on AI systems, Vision Transformers & on-device ML — medium.com/@ehtashamarif
- Member, COLAB NU (AI/DevOps community), 2024–2026
Skills
Tools of the trade
NLP & LLM / Agentic AI
Computer Vision
Machine Learning & Deep Learning
Backend & DevOps
Data & Big Data
Databases
Mobile & Languages
BS Computer Science
FAST NUCES
CGPA 3.25/4.0 · Dean's List 2025 (Top 10%) · Trilliet AI Hackathon 2025 Finalist