I'm a final-year Computer Science student who likes taking ML ideas from paper to working, deployable systems. My work sits at the intersection of computer vision, retrieval-augmented generation, and multimodal learning, backed by a full-stack background so I can ship what I build.
| 🎓 Education | BS Computer Science, UET Lahore (2023 – 2027) |
| 💼 Experience | ML Engineer Intern @ Medcare MSO · previously ML Intern @ NCAI |
| 🎯 Interested in | AI/ML research and applied AI |
| 💬 Ask me about | Computer vision · Medical AI · RAG systems · Full-stack development |
Three DOI-archived preprints on Zenodo, covering RAG robustness, medical VQA, and mental-health NLP.
|
RobustRAG Empirical study of corpus poisoning and query embedding noise in retrieval-augmented QA (BGE + FAISS, SQuAD 1.1). Query noise dominates over poisoning at the tested settings, and a common near-duplicate suppression defense shows a structural failure mode. |
|
MedInsight Does retrieval augmentation (CLIP + FAISS) improve a BLIP-2 vision-language model on medical VQA? Benchmarked on VQA-RAD and ROCOv2. |
|
AI-Based Depression Detection Fine-tuned DistilBERT + BiLSTM with attention for detecting depressive symptoms in social media text: 91.09% accuracy with interpretable predictions. |
| Languages |
|
| AI / ML |
|
| Full-Stack |
|
| Cloud & Services |
|
| Tools |
|
Current research · Label-efficient segmentation in histopathology
Semi-supervised melanoma and nuclei segmentation using a Mean Teacher framework, compared against a supervised U-Net baseline on the PUMA dataset under limited-label settings.
PyTorch U-Net Semi-Supervised Learning Medical AI
Retrieval-augmented medical image understanding · Preprint
Grounds a vision-language model in retrieved evidence (CLIP embeddings + FAISS) to test whether answer reliability improves on medical VQA.
PyTorch CLIP FAISS BLIP-2
🛡️ RobustRAG
RAG robustness and corpus poisoning · Preprint
Evaluates retrieval robustness under near-duplicate, contradictory, and irrelevant poisoning, plus a lightweight suppression defense, using BGE + FAISS on SQuAD 1.1.
BGE FAISS RAG Robustness
🕸️ GraphRAG
Adaptive hierarchical Graph-RAG
Extends Microsoft's GraphRAG with an adaptive retrieval router (local / global / hybrid / none), incremental graph updates without full re-indexing, and a self-verification layer that checks generated claims against retrieved evidence.
Neo4j Qdrant Streamlit LLM Orchestration
👁️ VisionTrack
Real-time object detection and tracking
Detection, tracking, and counting pipeline with configurable zones and counting logic.
YOLOv8 OpenCV Python
🔀 DevFlow
Project management platform
MERN-stack tool with real-time updates, Swagger API docs, Jest tests, and Docker-based deployment.
MongoDB Express React Node.js Socket.io Docker
Employee productivity monitoring platform
Activity tracking, OCR-based screenshot analysis, automated reporting via n8n, and a Supabase-backed analytics dashboard.
Flask Supabase EasyOCR n8n
🏆 GitHub Trophies
Microsoft Learn badges (4)
Intro to AI Concepts · Intro to Generative AI & Agents · Plan & Prepare AI Solutions on Azure · Intro to AI Speech Concepts
Building practical software systems with AI, one project at a time.