At Newmind.ai, I operate a production model-serving platform on Kubernetes, running more than 20 models across four clusters behind one inference API. I standardized GitOps deployment, rollout, and rollback for the platform.
I consolidated embedding and reranking models onto shared GPUs, using FP16 and time-slicing to sustain 105K inference requests a day at under 25% GPU utilization. I also introduced scale-to-zero for low-traffic models, reclaiming GPU capacity while holding cold starts at 12–15 seconds.
I automated GenAI model onboarding with a NATS-triggered control plane that generates Triton and KServe specifications and checks readiness. I deployed an LLM agent stack using MCP tool servers, a RAG agent, Neo4j, and Elasticsearch, and built dashboards and tracing for model observability.
Previously, at Gesund.ai, I contributed to the SDK for an AI as a Medical Device platform and cut application response time by 45%. At Vitamu, I shipped validation and drift-monitoring pipelines for an FDA-cleared clinical decision-support product; at Dogus Teknoloji, I built a real-time PPE detection system and predictive-maintenance models.
