Own model serving for multiple LLM/speech models on Modal. Build and maintain the APIs around those models. Create the feedback/eval loop to improve quality while meeting strict latency/cost SLOs. Responsibilities Host and scale real-time & batch inference on Modal (autoscaling, images/volumes/secrets). Operate a multi-model fleet (versioning, routing, canaries/blue-green, traffic shaping). Ship endpoints; auth, RBAC, quotas, rate limits, telemetry. Implement feedback pipelines, online A/B evals, and guardrails with actionable alerts. Drive performance: profiling, batching, quantization, KV-cache, runtime tuning. Establish observability and reliability (OTel, metrics/logs, SLOs, runbooks, on-call). CI/CD and IaC for reproducible builds and one-click rollbacks. Must-haves 5+ years in ML Ops/Platform/SRE with production LLM/ML serving. Strong Python; high-throughput async APIs (FastAPI/Starlette) and GitHub-based CI/CD. Deep experience with vLLM, TensorRT-LLM, Triton, or ONNX Runtime. Hands-on with Modal or equivalent GPU/k8s platform. Solid observability (OTel) and incident response/postmortems. Preferred ONNX export expertise (PyTorch→ONNX), quantized/dynamic graphs, custom ops. Safety/guardrails and constrained decoding. Systems perf (CUDA/Triton kernels) or Rust for hot paths; load/chaos testing.