Lovina Dmello

Lovina Dmello is a senior infrastructure software engineer on NVIDIA’s Deep Learning Libraries team, building systems that make production AI more reliable, observable, and secure. Her work spans accelerated inference, model-serving infrastructure, and practical safeguards for AI-assisted software development.

Dmello earned an engineering degree from the University of Mumbai and a master’s degree in computer science from the University of Georgia, where her graduate research on ransomware classification applied machine learning to the runtime behavior of Windows malware. She subsequently spent three years on Oracle Cloud Infrastructure and four years building backend systems for Apple Payments and Wallet before joining NVIDIA.

• Production AI security is infrastructure security. Dmello traces serious failures to unauthenticated services, excessive privileges, exposed credentials, flat networks, and unprotected model storage. Her analysis of production LLM security uses exposed Ray deployments to illustrate why least-privilege access, short-lived credentials, verified service identities, encryption, and network segmentation matter more immediately than speculative model attacks.

• Security controls must survive production economics. Authentication and input validation belong in every deployment; expensive workload isolation and runtime inspection should reflect actual risk, latency budgets, and throughput requirements. Dmello also identifies gaps between security, machine-learning, and operations teams as a recurring source of preventable vulnerabilities.

• Inference reliability begins before deployment. Her work on model-serving pipelines addresses failed model exports, unsupported operators, variable input dimensions, and dependency mismatches through continuous-integration checks, TensorRT optimization profiles, version tracking, and production-oriented profiling.

• Long-running model builds need operational visibility. Using TensorRT’s IProgressMonitor interface, Dmello demonstrates nested progress reporting and controlled cancellation in Python and C++, including integrations with developer tools, services, and agent runtimes.

• Coding assistants need safeguards outside the model. Her self-hosted coding-assistant architecture combines StarCoder2, NVIDIA NeMo Guardrails, dependency verification, commit-level traceability, and continuous-integration checks. She evaluates deployment quality through escaped defects, rollbacks, review latency, and incidents, while identifying prompt injection, retrieval-data poisoning, shared-GPU side channels, and model supply chains as additional security concerns.