We are building the MLOps backbone for enterprise generative AI, supporting both air-gapped private GPU clusters (for sensitive financial/healthcare data) and hyperscale cloud platforms (AWS SageMaker, Google Vertex AI, Azure ML). You will architect infrastructure that handles billions of inference tokens monthly across hybrid environment optimizing Llama 3/Mistral deployments on private A100 clusters while orchestrating GPT-4/Claude fine-tuning pipelines on managed cloud services. This is a high-visibility role requiring deep expertise in LLM serving engines, distributed GPU systems, and cloud-agnostic MLOps.

II. CORE RESPONSIBILITIES :

- Architect self-hosted inference clusters using vLLM, TGI (Text Generation Inference), and TensorRT-LLM on on-premise NVIDIA DGX systems and GPU racks, ensuring sub-100ms latency for 70B+ parameter models.

- Design parallel workflows on AWS SageMaker (Endpoints/Pipelines), Google Vertex AI (Prediction/Training), and Azure ML for elastic training workloads and managed foundation model APIs.

- Implement cloud-agnostic model deployment using Kubernetes (EKS/GKE/AKS) with portability across private data centers and cloud VPCs, ensuring zero vendor lock-in.

- Deploy multi-GPU inference parallelism (tensor + pipeline parallelism) for foundation models using Ray Serve, NVIDIA Triton, and custom FastAPI stacks.

- Optimize inference economics through quantization (AWQ/GPTQ/FP8), KV-cache optimization, and continuous batching - reducing per-token costs by 40%+.

- Build auto-scaling GPU node pools (Karpenter/Cluster Autoscaler) that respond to inference demand spikes within seconds.

- Implement RLHF (Reinforcement Learning from Human Feedback) infrastructure using DeepSpeed, LoRA/QLoRA fine-tuning pipelines, and distributed training orchestration.

- Design evaluation frameworks for LLMs : automated benchmarking (MMLU, HumanEval), A/B testing for model versions, and human-in-the-loop feedback systems.

- Manage vector database infrastructure (Pinecone, Weaviate, Milvus, pgvector) for RAG systems spanning private and cloud environments.

- Build CI/CD for ML using GitOps (ArgoCD/Flux) with model versioning (MLflow/DVC), automated testing for data drift, and canary deployments for model updates.

- Implement feature stores (Feast/Tecton) and experiment tracking (Weights & Biases/MLflow) supporting both cloud and on-premise data lakes.

- Create observability stacks for LLMs : token-level latency tracking, GPU memory saturation alerts, and cost-per-inference dashboards using Prometheus/Grafana/CloudWatch.

- Manage secrets, model encryption at rest (HashiCorp Vault), and network policies (Istio/Linkerd) for multi-tenant model serving.

III. ESSENTIAL QUALIFICATIONS & EXPERIENCE :

Educational Qualifications :

- Bachelor's degree (B.E./B.Tech) in Computer Science, Engineering, Mathematics, or related technical field from a recognized university. Graduates from IITs, NITs, BITS, IIIT, or top-tier engineering institutions preferred.

- Master's degree (M.Tech/MS) in Machine Learning, Computer Science, Artificial Intelligence, or related field desirable but not mandatory.

- Relevant professional certifications in cloud platforms (AWS/Azure/GCP) and Kubernetes (CKA/CKAD) highly desirable.

Experience Requirements :

- Minimum 5 - 9 years of hands-on experience in production ML infrastructure engineering, with at least 2 years dedicated to large-scale model deployment and MLOps.

- Demonstrable track record of deploying and maintaining 70B+ parameter models in production environments (are preferred).

- Proven experience managing both on-premise GPU clusters (NVIDIA DGX, A100/H100) and cloud-based ML platforms (AWS SageMaker, Google Vertex AI, or Azure ML).

IV. TECHNICAL COMPETENCIES REQUIRED :

Infrastructure & Systems :

- Expert-level proficiency in Kubernetes (GPU operators, taints/tolerations, multi-tenancy) across both on-premise (Rancher/OpenShift) and cloud (EKS/GKE/AKS) environments.

- Deep expertise in LLM serving engines : Proven hands-on experience with vLLM, TGI (Text Generation Inference), or TensorRT-LLM in production settings.

- Professional-level certification or equivalent experience in AWS SageMaker, Google Vertex AI, or Azure ML - including model registry, endpoints, and pipeline orchestration.

- Strong understanding of NVIDIA Hopper/Ampere architectures, NVLink/InfiniBand networking, and CUDA optimization.

- CUDA kernel optimization, custom inference kernels, or TritonML server extensions.

- Infrastructure as Code : Terraform, Helm, Kustomize for reproducible GPU cluster provisioning.

Machine Learning & Distributed Systems :

- Expert-level Python programming with PyTorch/TensorFlow.

- Distributed training frameworks : DeepSpeed, Horovod, PyTorch DDP/FSDP.

- LLM Stack : LangChain, LlamaIndex, Hugging Face Transformers, and agentic workflow orchestration.

- Data Engineering : Apache Spark, Airflow, and feature engineering at scale (terabyte+ datasets).

- Database Systems : Vector databases (Pinecone, Weaviate, Milvus, pgvector) and feature stores (Feast/Tecton).

DevOps & Observability :

- CI/CD for ML : GitOps (ArgoCD/Flux), model versioning (MLflow/DVC), and automated testing.

- Observability : Prometheus, Grafana, ELK stack, and cloud-native monitoring (CloudWatch/Stackdriver/Azure Monitor).

- Security : HashiCorp Vault, Istio/Linkerd service mesh, network policies, and secrets management