Haseeb Mustafa

Machine Learning Engineer

Stamatis • Freelance

Apr 2026 - Aug 2026 • 4 mos

Built and deployed a production medical-imaging AI platform end to end — from raw datasets to a live, authenticated web product on AWS. Designed and trained deep-learning models across four anatomical domains (brain, heart, chest, spine) spanning 3D MRI, 3D cine cardiac MR, 2D X-ray, and 3D CT, all served behind a single FastAPI backend and an interactive diagnostic dashboard. Highlights: - Architected a modular 4-model platform on a shared 3D "mask-then-read" backbone (73.9M params), each organ shipped as an isolated plug-in module (router + loader + status) so new body parts drop in without touching existing code. - Trained the brain-tumour (3D MRI) and cardiac (3D cine MR) models to clinical-grade held-out metrics — brain ~94% accuracy / 95% sensitivity, heart ~92% accuracy on a 5-class cardiomyopathy task — using subject-grouped cross-validation to prevent patient-level data leakage. - Cut inference compute ~73% via a probabilistic masking pipeline, enabling full 3D volumetric inference on CPU-only production hardware. - Built a self-supervised (3D masked autoencoder) pretraining pipeline for label-scarce domains and integrated automatic CT segmentation (TotalSegmentator / nnU-Net) so users can upload raw scans with no manual annotation. - Designed the full frontend — organ switcher, drag-and-drop upload, HIPAA flow, animated inference progress, overlay rendering, and interactive 3D volume viewing (NiiVue). - Deployed and maintained the live system on EC2 (Nginx + systemd + Google SSO) with safe staged deploys, health checks, and auto-rollback; cut cloud costs ~80% via infra right-sizing and a zero-downtime migration. Stack: PyTorch, 3D CNNs, self-supervised learning, DenseNet, nnU-Net, NiBabel, SimpleITK, pydicom, scikit-image, OpenCV, NumPy/SciPy/pandas, scikit-learn, FastAPI, Uvicorn, OAuth/SSO, JavaScript, NiiVue, AWS EC2/EBS, nginx, systemd, Linux.

Machine Learning Engineer

Mulligan LLC • Temps partiel

Aug 2025 - Jun 2026 • 10 mos

1. Built MSA Scout, an AI-powered property investment platform. Developed full-stack system with market data analysis, smart filtering, and dashboards to help investors find high-growth real estate opportunities. 2. Build Predictive Models for bond holdings capacity for USA banks using FDIC call reports data and increase revenue by 60% 3. Build ML model using Census, BLS data to predict top 10 MSA in USA to invest in property sectors