Ofertas › Fireworks
Fireworks · San Mateo · Remoto · $175K – $220K • Offers Equity
ABOUT US:
Fireworks is the platform for specialized intelligence, enabling companies to build, train, and serve AI models tailored to their own data, workflows, and products. Founded by the team behind PyTorch and backed by AMD, Atreides, Benchmark Capital, Index Ventures, Lightspeed, NVIDIA, Sequoia Capital, and TCV, Fireworks powers production AI with hundreds of state-of-the-art open models across text, image, embedding, audio, and multimodal workloads. Today, Fireworks is a Series D company valued at $17.5 billion, bringing together an ambitious, collaborative team that's building the future of enterprise AI.
The Role:
Make data a solved problem for every GPU workload at Fireworks. We own how data is stored, cached and moved across clouds, so that training and inference teams never have to think about where their data lives or how it gets to them, and no GPU ever waits on it.
In practice, that means building a unified, high-performance data layer that spans parallel file systems, object storage, distributed caches and key-value stores across multiple clouds. Datasets, checkpoints, model weights and KV cache should be in the right place at the right time, at whatever scale the GPUs need. Training jobs start without a data-staging step, inference replicas cold-start without waiting on weight downloads, and the system handles placement, replication and movement on its own.
Key Technical Archetypes:
- We expect strong candidates to have gone deep in at least one of these areas and to be comfortable in the others:
- The Storage Systems Builder: Has designed and shipped a storage system: a distributed file system, a blob/object store, a KV store, or a storage engine. They can explain the tradeoffs they made in consistency models, replication, erasure coding, metadata scaling, compaction and failure recovery, and why they made them.
- The Caching Architect: Has built a distributed or tiered cache across memory, NVMe and remote storage. They think in hit rates, eviction policies, cache coherence, prefetching and tail latency. Bonus if they've worked on caching for model weights, datasets, checkpoints, or LLM KV cache offload.
- The Linux Power User: Knows the system under the hood: kernel I/O paths, POSIX semantics, page cache behavior, io_uring/SPDK, and OS-level tuning (sysctl, I/O schedulers, NUMA).
- The Network-As-Code Expert: Understands RDMA, RoCE v2, InfiniBand and GPUDirect Storage from a software and configuration perspective, because