Operations Engineer, HPC Networking

fal is the generative media ecosystem powering the next generation of AI products. We build the infrastructure, tools, and model access that teams need to move from idea to production, and do it at scale without compromise. For developers and enterprises, fal is the foundation that makes generative media not just possible, but practical: a unified platform where high-performance inference, orchestration, and observability come together to unlock new categories of AI-native products.

As generative media reshapes industries across a market projected to grow by hundreds of billions over the next decade, fal is becoming the ecosystem that ambitious teams build on.

We're hiring an Operations Engineer for HPC Networking to keep our InfiniBand and Ethernet fabrics healthy as we scale.

This is a hands-on role. You'll bring up new fabrics alongside DC ops, monitor the ones in production, and chase down the weird stuff: link flaps, congestion, NCCL stalls, firmware bugs that only show up at scale. 

• Operated InfiniBand fabrics in production: subnet manager, routing, partitioning, monitoring.
• Debugged the full stack: cables, transceivers, switch firmware, HCAs, drivers, NCCL.
• Brought up new fabrics from cable pull through validation.
• Scripted your way through repetitive operational work (bash, python, go, whatever).
• Nice to have: Ethernet RoCE, Spectrum-X, or large-scale GPU cluster networking.

• Detail-oriented. Cable plant hygiene is a personality trait.
• Calm under fire. A fabric incident during a customer training run doesn't rattle you.
• You read vendor release notes for fun, or at least out of self-defense.
• You'd rather find the root cause than reboot the switch.

• →Monitor health and performance of InfiniBand and Ethernet fabrics: switches, HCAs, transceivers, links.
• →Investigate and resolve fabric issues: connectivity, congestion, performance regressions.
• →Support fabric bring-up alongside DC ops and customer-facing teams.
• →Run maintenance and upgrades on switches and control plane components.
• →Partner with cluster ops on cross-domain incidents where the line between compute and network is blurry.
• →Improve the tooling and runbooks so the next incident resolves faster than the last.