Systems & Architecture Archive.

— Architected a 4-component staged decisioning pipeline routing 284K+ transactions through distinct uncertainty layers: V1 (abstention-aware XGBoost ensemble router with Isolation Forest novelty detection), V2 (calibrated SVM second-opinion for epistemic uncertainty resolution), V3 (Dempster-Shafer belief fusion across 3 independent evidence sources), and V4 (SHAP-backed structured deferral replacing human queues).

— Achieved 100% automation rate (101 → 92 → 72 → 0 human review cases across stages), 86.73% fraud recall, 100% DECLINE precision (zero false blocks on legitimate transactions); deployed FastAPI REST API with sub-100ms inference.

— Built a scalable platform for teacher training & evaluation with async FastAPI backend, PostgreSQL with pgvector for semantic plagiarism detection, and Celery + Redis task queue for background AI inference and OCR jobs.

— Designed RBAC system (Super Admin / Institution Admin / Teacher) with RS256 JWT auth, PgBouncer connection pooling, SQLAlchemy 2.0 async ORM; containerized full stack with Docker Compose.

— Engineered a 5-stage cryptographic pipeline for multimodal medical data (DICOM, EHR, reports) achieving 120+ MB/s encryption and 130+ MB/s decryption with <0.1% storage overhead.

— Custom binary container format with cross-modal HMAC binding to prevent splicing attacks; RSA-OAEP digital envelope for secure key transit.

— Built a Legal Metrology label compliance scanner combining dual-engine OCR with OpenCV preprocessing to extract and validate mandated label fields.

— Designed an explainable rule engine where every pass/fail verdict is traceable to a specific regulation clause, served through an async Celery/Redis pipeline.