AI Infrastructure Engineer
MercorSan FranciscoOn-Site
Needs a free Superjobs account. You'll finish on Mercor's Ashby page.
Skill score
Summary
The AI Infrastructure Engineer owns Mercor's GPU accelerator platform end to end, building multi-cloud clusters, schedulers, LLM serving, health-checks, and cost attribution. The role suits experienced software engineers who want to write production code and own distributed systems and on-call work rather than operations-only tasks.
- Function
- Development
- Estimated salary
- Est. $138k–208k/yr
- Estimated for a mid-level role, ~5 yrs experience, United States. Not published by the employer.
- Work mode and location
- On-Site, San Francisco
- Experience
- 5+ yrs
- Employment type
- Full-time
Job description
Overview
All of Mercor’s AI work, from training and serving models to building agents, runs on GPUs. Your job is to make that compute fast, fault-tolerant and fully utilized.
You’ll own the accelerator platform end to end. That means the clusters and schedulers across AWS, GCP, neoclouds and on-prem, the training and LLM serving on top, and the observability and cost data that show where every GPU-hour goes.
This is a software engineering role, not an ops role. You’ll write production code every day, own systems from design through on-call, and build the platform that sets how fast Mercor can move.
Roles and responsibilities
- Stand up a multi-cloud GPU fleet spanning hyperscalers, neoclouds and on-prem, orchestrated through custom Kubernetes and Slurm operators so researchers never have to think about where their job lands.
- Build a scheduling layer with gang scheduling, bin-packing and fair-share quotas that measurably raises accelerator utilization across teams.
- Ship a high-throughput vLLM serving stack with prefill/decode disaggregation, KV cache offload and FP8/FP4 quantization. Back it with a benchmarking and traffic-replay harness that tracks TTFT, TPOT and latency percentiles against per-config baselines.
- Build a self-healing health-check and remediation system that rides out fault storms across hardware, network and software without paging a human.
- Build an end-to-end cost-attribution pipeline. It ingests utilization, occupancy and billing data into the warehouse, attributes every GPU-hour to a team and workload, and drives capacity planning.
- Build internal tools and dashboards that researchers and non-engineers use every day.
Responsibilities
- Design, deploy and operate GPU clusters across AWS, GCP, and other hyperscalers/neoclouds/on-premise data centers, including custom Kubernetes and Slurm operators and the metadata services behind them.
- Build scheduling that keeps accelerators busy: gang scheduling for distributed jobs, bin-packing to reduce fragmentation, and quotas and fair share for multi-tenant teams.
- Run distributed training infrastructure on PyTorch Distributed, Ray, or Megatron, with fast checkpointing and model loading.
- Deploy and tune high-throughput LLM serving with vLLM or similar: tensor and model/expert parallelism, prefill/decode disaggregation, KV cache offload, FP8/FP4 quantization.
- Build benchmarking and traffic-replay tools that measure TTFT, TPOT and latency percentiles, and set per-config performance baselines.
- Build health checks across hardware, network and software layers, and automated remediation that handles fault storms without human intervention.
- Instrument fleet utilization with DCGM, Prometheus and Grafana. Write recording rules and alerts that teams rely on.
- Ingest utilization, occupancy and billing data into a warehouse, attribute spend to teams and workloads, and support capacity planning.
- Drive autoscaling, rightsizing and idle-capacity recovery programs with measurable cost results.
- Write clear runbooks and documentation, and explain infrastructure tradeoffs to researchers, finance and leadership.
- Build internal tools or dashboards that researchers and non-engineers use daily.
Skills required
Top skills
- Python
- Go or C++
- Shell scripting
- Prometheus and Grafana
All skills (8)
- Kubernetes
- GPU workloads at scale
- Slurm
- RDMA, InfiniBand, NCCL
- PyTorch Distributed
- Ray
- vLLM
- Megatron-LM
Qualifications
- Have 5+ years building and operating production distributed systems or ML infrastructure
- Write production-quality Python plus Go or C++, and are fluent in shell scripting for system administration.
- Have deep hands-on Kubernetes and container experience, plus working knowledge of Slurm or another HPC resource manager.
- Have run GPU workloads at scale, including scheduling, orchestration and troubleshooting.
- Have hands-on experience with at least one distributed training or serving stack: PyTorch Distributed, Ray, vLLM or Megatron-LM.
- Have deep experience with at least one major cloud provider.
- Know observability tools like DataDog, Prometheus, PromQL and Grafana, and are comfortable on-call for critical systems.
- Own projects end to end and gather your own requirements with little direction.
Nice to have
- LLM inference optimization: KV cache systems, disaggregated serving, quantization, speculative decoding
- High-performance networking: RDMA, InfiniBand, NCCL, NVLink
- GPU profiling with Nsight or similar, and CUDA or Triton kernel work
- Accelerators beyond NVIDIA: AMD MI-series, TPU, Trainium
- Capacity planning, billing normalization or cost attribution across providers
- Automated fault diagnosis, log mining or AIOps at fleet scale
- Configuration management (Ansible, Puppet) and secure or air-gapped deployments
Other details
- Generous equity grant vested over 4 years
- Health, dental, and vision insurance
- 401(k) with company match
- Unlimited paid vacation and sick days
- Parental leave for birthing and non-birthing parents
- Free mental health services
- $200 monthly wellness reimbursement
- $1.5K monthly meal stipend
- $200 monthly laundry reimbursement
- Free Equinox membership
- Up to $15K relocation bonus
Company info
About Mercor
Mercor’s mission is to organize human intelligence for the AI economy. We build the infrastructure that enables frontier AI labs and enterprises to train, evaluate, and improve AI on real-world professional work. More than five million experts have joined our network, and we pay out over $5 million daily for