class EhtashamArif:
    def __init__(self):
        self.name        = "Ehtasham Arif"
        self.role        = "AI / ML Engineer"
        self.university  = "FAST NUCES — Dean's List 2025 🏆"
        self.semester    = "6th Semester, BS Computer Science (2023–2027)"
        self.location    = "Islamabad, Pakistan 🇵🇰"

    @property
    def currently(self):
        return [
            "🤖  Building LLM & RAG pipelines and agentic workflows (LangChain, LangGraph, CrewAI) — AI/ML Intern @ Xandec",
            "🔭 Researching Vision Transformers & cross-hospital medical AI generalization",
            "🛡️  Building deepfake detection systems with EfficientNet + ELA forensics",
            "📱  Developing on-device AI photo editing apps with MIRNet + TFLite",
            "🌱  Deepening knowledge in LLMs, MLOps, and big data pipelines",
            "🤝  Teaching Assistant — Programming Fundamentals & DB Systems @ FAST NUCES",
            "✍️   Writing about AI/ML at blog.ehtashamarif.com",
        ]

    def contact(self):
        return "p230567@pwr.nu.edu.pk"

Category	Technologies
Core ML / DL
Transformers & Vision
Computer Vision
LLMs & Agentic AI
NLP & Attention
Research & Kaggle
Data Science
Graph & Network
Big Data (Learning)

Category	Technologies
Web Dev
Mobile
Databases
DevOps
Languages

PyTorch · SAM 3D · CatVTON · Flutter · Computer Vision · Fine-Tuning

A deep learning computer vision pipeline for body measurement estimation and virtual garment try-on, built as part of applied AI work at Xandec.

• Fine-tuned SAM 3D for body segmentation, extracting precise body contours from images to drive measurement estimation
• Fine-tuned CatVTON for realistic garment fitting, rendering clothing onto a user's body with accurate drape and fit
• Delivered the full pipeline through a companion Flutter mobile app for an end-to-end try-on experience
• Combined classical CV preprocessing with deep generative fitting for a production-oriented virtual try-on flow

PyTorch · Vision Transformer (ViT-B/16) · timm · Kaggle GPU · Medical AI · Custom Loss Design

The Problem: An AI trained at Hospital A fails at Hospital B — not because diseases changed, but because the model learned scanner artifacts instead of actual pathology. This is the cross-hospital generalization gap — a critical barrier to real-world clinical deployment.

The Solution: A novel Clinical Attention Consistency (CAC) Loss that supervises ViT internal attention maps with 880 radiologist-drawn bounding box annotations, forcing the model to attend where doctors attend.

Standard Fine-tuning  →  Model looks at random artifacts and bright spots
CAC Loss (Our Method) →  Model forced to focus on clinically relevant disease regions ✅
Total Loss = BCE Loss + 0.5 × CAC Loss

• Fine-tuned ViT-B/16 with partial unfreezing (blocks 8–11) + custom weighted BCE for extreme class imbalance (up to 637:1)
• Extracted CLS→patch attention from the final transformer block to compute spatial supervision signal
• Full research paper written to publication standard, published open-access on Zenodo

📄 https://doi.org/10.5281/zenodo.19809678

PyTorch · EfficientNet-B0 · FastAPI · React 19 · OpenCV · ELA Digital Forensics

Professional full-stack AI forgery detection combining deep learning with classical digital forensics.

• Hybrid dual-pipeline: EfficientNet-B0 deep learning backbone + Error Level Analysis (ELA) for JPEG compression artifact detection
• Face-centric analysis: OpenCV Haar Cascade isolates and crops faces for targeted GAN/Diffusion artifact classification
• Weighted consensus verdict: Face model (60%) + Global model (25%) + ELA forensics (15%) → classifies as REAL / FAKE / UNCERTAIN
• Full-stack deployment: FastAPI backend with Uvicorn + React 19 frontend delivering real-time confidence scores and forensic breakdowns
• Dual inference runs in parallel on full image and isolated face region to catch both global scene inconsistencies and local facial artifacts

Flutter · C++17 · TFLite · MIRNet · dart:ffi · CMake · Provider

Professional AI-powered photo editor that runs entirely on-device — no internet, no cloud, no data leaving your phone.

• On-device MIRNet AI enhances low-light images with a ~27MB TFLite model via tflite_flutter — zero server dependency
• C++17 native processing core accessed through dart:ffi for maximum speed and memory efficiency; parallel isolate-based inference keeps UI at 60fps
• 21+ pro editing tools: AI Smart Enhance, 8 style filters (Pencil Art, Anime, Comic Book, Retro Film, 3D Relief…), manual adjustments (brightness, contrast, saturation, gamma, exposure, sharpness), face retouching, geometry tools
• Non-destructive editing pipeline: Full undo/redo history with memory-efficient JPEG compression at every step
• Auto-detects hardware acceleration capabilities; gracefully falls back to native C++ processor when AI acceleration is unavailable

Python · MediaPipe 0.10 · OpenCV · HTML5 Canvas · Modular CLI

A unified, modular suite of production-ready computer vision modules for real-time human body tracking and gesture interaction.

• Face Mesh: 468 facial landmark tracking at sub-pixel accuracy — suitable for AR overlays and expression analysis
• Multi-Hand Tracking: Simultaneous detection of multiple hands with 21 keypoints each; built-in gesture recognition pipeline
• Holistic Pose Detection: Full-body skeleton tracking using 33 body keypoints for fitness / motion capture workflows
• AirDraw Web App: HTML5 canvas application — draw in digital space via hand gestures tracked through webcam, no hardware required
• Smart run.py launcher: Auto-checks and installs all dependencies, presents a CLI module selector — zero-friction onboarding

Node.js · Nginx · Docker · Linux VPS · SSL/TLS

Production-deployed full-stack real-time image enhancement platform on a Linux VPS. Configured Nginx reverse proxy, SSL termination, and containerised deployment. Optimised server-side processing pipeline to reduce response latency. Live at sketchey.com.

┌─────────────────────────────────────────────────────────────────────┐
│                                                                     │
│  🎓  BSc Computer Science          FAST NUCES · 2023 – 2027        │
│  🏆  Dean's List 2025              Top 10% of class                │
│  🤖  AI Hackathon Finalist         Trilliet AI Hackathon 2025      │
│         Top teams from 50+ competing                               │
│  🔬  AI Research Participant       NuSYS Lab, FAST NUCES           │
│  📄  Published Researcher          Zenodo · 2025                   │
│  📚  Teaching Assistant            Prog. Fundamentals & DB Systems  │
│  👥  Core Member                   COLAB NU Tech Community         │
│                                                                     │
└─────────────────────────────────────────────────────────────────────┘

---