SlipstreamJobs tracks this role from the company's public career site. Apply directly on the employer's site.
Fireworks is a Series D AI infrastructure platform (valued at $17.5B, backed by NVIDIA, Sequoia, Benchmark, and others) that enables companies to build, train, and serve AI models tailored to their data and workflows. The company is founded by the PyTorch team and powers production AI with hundreds of state-of-the-art open models across text, image, embedding, audio, and multimodal workloads.
You will own the data infrastructure layer that makes data a solved problem for every GPU workload at Fireworks. This means designing and building a unified, high-performance data layer spanning parallel file systems, object storage, distributed caches, and key-value stores across multiple clouds. Your work ensures that datasets, checkpoints, model weights, and KV cache are in the right place at the right time, at whatever scale GPUs demand. Training jobs start without data-staging steps, inference replicas cold-start without waiting on weight downloads, and the system handles placement, replication, and movement autonomously.
The role targets one of four technical archetypes (though depth in one area is valued over breadth):
1. **Storage Systems Builder**: Design and ship distributed file systems, blob/object stores, KV stores, or storage engines. You explain tradeoffs in consistency models, replication, erasure coding, metadata scaling, compaction, and failure recovery.
2. **Caching Architect**: Build distributed or tiered caches across memory, NVMe, and remote storage. You think in hit rates, eviction policies, cache coherence, prefetching, and tail latency—ideally with experience caching model weights, datasets, checkpoints, or LLM KV cache offload.
3. **Linux Power User**: Deep knowledge of kernel I/O paths, POSIX semantics, page cache behavior, io_uring/SPDK, and OS-level tuning (sysctl, I/O schedulers, NUMA).
4. **Network-As-Code Expert**: Understand RDMA, RoCE v2, InfiniBand, and GPUDirect Storage from software and configuration angles, since neoclouds often require manual plumbing.
You will collaborate with world-class engineers and AI researchers in a fast-growing, flat organization where your work directly shapes the future of AI infrastructure.
**Requirements:**
- Bachelor's degree in Computer Science, Engineering, or related technical field (or equivalent practical experience).
- 5+ years of experience designing, building, and operating storage, caching, or data infrastructure in production at scale.
- Strong software development skills in C/C++, Rust, or Go, with experience writing performance-critical data path code. Python or Go for automation.
- Deep understanding of distributed storage fundamentals: consistency models, replication and consensus (Raft/Paxos), metadata scaling, caching, and failure recovery.
- Hands-on experience with Kubernetes, including CSI, to mount high-performance volumes into ephemeral GPU pods.
- Experience deploying and operating storage clusters (WEKA, VAST, Lustre, Ceph, MinIO) across clouds using infrastructure-as-code tools (Terraform, Pulumi).
- Proficiency with performance benchmarking and profiling tools (fio, iperf3, NCCL tests, perf, eBPF) to identify and resolve bottlenecks across kernel, network, cache, and storage backend.
**Preferred qualifications:**
- Core contributor or owner of a storage, caching, or KV system (in-house or open source: Ceph, Alluxio, JuiceFS, RocksDB, FoundationDB, TiKV, Redis/Dragonfly, SeaweedFS, MinIO).
- Experience with high-performance computing or large-scale AI clusters (Slurm, Kubernetes GPU pools) and familiarity with AI I/O patterns (checkpoint bursts, random dataset reads, model-weight loading for inference cold starts).
- Familiarity with Linux internals: kernel I/O paths, POSIX semantics, page cache behavior, io_uring/SPDK, OS-level tuning.
- Familiarity with high-performance networking (RDMA, RoCE v2, InfiniBand) and GPUDirect Storage.
- Experience moving, replicating, or caching petabyte-scale datasets across clouds (AWS, GCP, specialized neoclouds) with attention to egress cost and consistency.
- Preference for software-defined storage over proprietary hardware appliances.
