Varun R.

ML engineer who builds LLM and production ML systems, plus the Python backends and infra they run on. That mix is my edge: my models don't just sit in a notebook; they run fast, cheaply, and reliably in production. 5 years in production ML. CKA certified. 49 projects on Upwork. What I build: Python backends and APIs: FastAPI, async services, REST and OpenAI-compatible endpoints, Redis for caching and feature stores, built to hold a p95 target in a live request path (I cut one serving path from ~25s to 4-5s under concurrent load). Platform and infra: containerized deployment across 50+ Kubernetes clusters, Terraform IaC, ArgoCD/Helm GitOps, and Prometheus/Grafana monitoring, hosted or in your own cloud. Cost-efficient GPU and inference: self-host open models (Qwen, testing MiniMax/GLM) to replace expensive APIs, and two-layer autoscaling (KEDA + Karpenter) that cut idle GPU spend ~50%. LLM and RAG systems: model serving on vLLM, RAG pipelines with monitoring on retrieval quality, and structured/guided decoding so outputs follow a fixed schema your backend can consume. Local and edge inference: quantized GGUF models profiled and tuned to run on CPU and constrained hardware, not just big GPUs. Stack: Python, FastAPI, PyTorch, vLLM, LiteLLM, RAG, MLflow, Kubernetes, AWS/SageMaker, Terraform, ArgoCD/Helm, Docker. Need it built hands-on or designed right? I work in your stack, send clear updates, and care about systems that keep running after I hand them off. Tell me what you're building and I'll give you an honest take.