Original listing text, shown exactly as published by the company.
What You Will Be Doing
- Own the Ray ecosystem end-to-end: manage KubeRay on GKE, tune Ray Core Task/Actor scheduling, operate the Plasma distributed object store, and configure Ray Data for GPU-direct streaming from GCS/S3
- Operate distributed training with Ray Train: configure TorchTrainer + DDP/NCCL for multi-node H100 clusters, manage checkpoint lifecycle, implement spot-preemption recovery, and integrate warm-start fine-tuning for retrain pipelines
- Build and operate the LLM inference mesh with Ray Serve: compose vLLM (PagedAttention), SGLang (RadixAttention), and NVIDIA Triton (TensorRT/ONNX) as a unified deployment graph with Plasma zero-copy memory sharing
- Optimise inference performance: configure fractional GPU allocation, enable continuous batching, implement per-engine autoscaling based on request queue depth, and tune KV-cache block sizes
- Design and operate the model routing layer: capability-based, version-based, and tenant-based routing with cost-aware fallback between self-hosted SLMs and cloud LLMs
- Build RL training infrastructure: define Flyte workflows for RL pipelines (rollout, reward shaping, policy update, evaluation), integrate Ray RLlib or custom PPO/GRPO loops with Ray Train, and manage replay buffer persistence on GCS
- Operate the full model promotion lifecycle: quality gate → integration tests → load tests (k6) → shadow mode → A/B gate → canary (10%→100%) with golden-signal auto-rollback
- Operate the retrain pipeline: drift detection triggers, warm-start retraining, relative quality gates (V2 >= V1 − 2%), and automated Flyte DAG through to canary
- Integrate RAG retrieval into the inference mesh: vector similarity search, context assembly, and prompt construction before LLM inference
What You Bring
- Experience in ML engineering with time in an ML platform or MLOps role
- Production Ray depth: Ray Train, Serve, Core, and Data — debugged real production failures including NCCL timeouts, Plasma OOM, and Serve autoscaling lag
- LLM serving engines: hands-on with vLLM, SGLang, or NVIDIA Triton — PagedAttention, prefix caching, and continuous batching tuned for latency/throughput targets
- Distributed training: DDP, FSDP, NCCL collectives, gradient checkpointing, and mixed precision (BF16/FP8)
- RL working knowledge: PPO, policy gradient, or RLHF — able to translate an algorithm into distributed compute primitives
- Model lifecycle operations: MLflow registry, shadow/A/B/canary patterns, and auto-
rollback on golden signal degradation
- Vector databases: Pgvector or Qdrant — ANN index strategies, embedding upsert, and query latency tuning under inference load
- Strong Python and PyTorch; Flyte or equivalent ML orchestrator
- Quantization (nice to have): INT8/INT4/FP8 post-training quantization (GPTQ, AWQ, or bitsandbytes)
- Bachelor's degree in Computer Science, Engineering, or a related field, or equivalent
practical experience or equivalent military experience
We offer you a competitive total rewards package, learning and tremendous opportunities to grow and advance in your career. At Saviynt, it is not typical for an individual to be hired at or near the top of the range for their role and final compensation decisions are dependent on many factors including, but not limited to location; skill sets; experience and training; licensure and certifications; and other relevant business and organizational needs.
You may also be eligible to participate in a Saviynt discretionary bonus plan, subject to the rules governing the program, whereby an award, if any, depends on various factors, including, without limitation, individual and organizational performance.