Software Engineer & AI/ML Engineer | final year at FAST NUCES Karachi. Building systems that hold up beyond the notebook, from computer vision pipelines to cloud-native backends and compliance-aware architecture.
Most ML projects look great in a Jupyter notebook. They fall apart the moment they touch the real world. I've been obsessed with one question since my second year at FAST NUCES: what does it actually take to build AI that holds up in production?
That obsession led me to build a Pakistan Sign Language Translator — an encoder-decoder deep learning pipeline built for a community of 70 million people underserved by technology. Not a demo. A working pipeline with real deployment constraints and real accuracy targets.
I work across the full stack: computer vision pipelines, cloud-native serverless backends, MLOps infrastructure, and compliance-aware data architecture. I don't just build features — I build systems that don't break when the assumptions do.
End-to-end AI pipeline translating Pakistan Sign Language into text. Encoder-decoder deep learning architecture, MediaPipe hand landmark extraction from raw video, and post-processing to generate grammatically coherent Urdu and English output. Built for 70 million people underserved by mainstream technology.
Serverless organ donor-receiver platform on AWS with automated matching logic, real-time SES notifications, and state-machine-enforced MySQL schema for donor-receiver lifecycle integrity.
Role-stratified HMS with SHA-256 and AES patient data anonymization, MySQL audit triggers on every CRUD operation, and runtime UI adaptation based on session role.
WebSocket chat with session-based auth and FTP file transfer. Network fault simulation — packet delay, loss, and corruption — to validate protocol reliability under real-world conditions.
Production-grade leave management microservice solving the dual-write inconsistency between an internal platform and an external HCM system (Workday/SAP). Handles race conditions via optimistic locking, HCM unreliability via exponential backoff with jitter, and duplicate submissions via idempotency keys with a full audit trail on every state change.
End-to-end ML pipeline that automates bug triage across 500,000 frontend UI/UX bug records. Benchmarked 7 classification algorithms across two targets — severity (4 classes) and priority (3 classes). Decision Tree achieved 88.71% accuracy on severity; XGBoost hit 89.63% on priority. Features selected via mutual information scoring, with a 15-step data cleaning pipeline from raw Kaggle data to production-ready serialized models.
Full-stack music streaming web app powered by the Spotify API and Firebase. Features Google and email authentication, real-time search across tracks, artists and albums, a 30-second preview player with shuffle and repeat, custom playlist management, and user profiles with photo upload — all persisted across sessions.
Open to Software Engineer and ML Engineer roles — remote or Karachi-based. If you're building something that matters, I want to hear about it.