Noor Aftab

Lead the design and operation of a distributed AI platform for KYC, identity verification, and document intelligence, processing 100,000+ verification requests per day across 60+ Kubernetes clusters.

• Architected end-to-end MLOps pipelines — data ingestion, model inference (OCR, computer vision, LLM extraction), evaluation, and monitoring — with confidence-based fallback and validation mechanisms
• Re-architected LLM inference using vLLM-based shared GPU services with continuous batching and optimized VRAM allocation, reducing GPU consumption from 60+ GPUs to 3–5 shared GPUs
• Achieved 60% infrastructure cost reduction, lowering monthly GPU/compute expenses from USD 20,000 to USD 7,500 through GPU consolidation and workload optimization
• Optimized end-to-end inference latency from ~10s to ~4s (60% reduction) by tuning model execution paths with TensorRT and ONNX Runtime
• Designed and maintained 100+ production microservices using FastAPI, deployed with Docker on Kubernetes via Terraform-managed infrastructure with automated CI/CD
• Implemented production observability using Elastic Stack and Grafana dashboards — request tracing, structured logging, performance metrics, and alerting
• Deployed and tuned models via NVIDIA Triton Inference Server and vLLM for production-grade latency and throughput across vision and LLM workloads
• Built agentic LLM tooling on Vertex AI that automates deployment orchestration, merge-request code review, and production latency investigation across multi-environment clusters

• YOLO
• Mask R-CNN
• PaddleOCR
• ONNX Runtime
• vLLM
• TensorRT
• NVIDIA Triton
• Vertex AI
• Gemini
• GPT
• Claude
• LLaMA
• Qwen
• Python
• FastAPI
• Cython
• Kubernetes
• Docker
• Terraform
• CI/CD
• Elasticsearch
• Grafana
• MongoDB
• PostgreSQL