Performance Engineer, Containers/Serverless

Verda is a technology company building the next generation of cloud infrastructure for AI. We operate GPU clusters across Europe,

the US, and Asia, and we run some of the most demanding AI/ML workloads in production today — from frontier-model training to

latency-sensitive inference at scale. We're a low-hierarchy team that ships pragmatically and gives engineers real ownership of

the systems they build.

Running ML/AI workloads in containers and serverless environments looks simple on a slide and is anything but in production.

Between the object store and the GPU sit a dozen layers — image pulls, network filesystems, page cache, model loaders, runtime

initialization — and each one quietly contributes to how long a workload takes to become useful and how fast it runs once it is.

We're hiring a Performance Engineer to own that surface for Verda's container and serverless GPU platforms. You'll characterize

where time and throughput actually go across the stack, and turn that understanding into concrete platform improvements.

Cold-start latency is one of the more visible expressions of the problem — a 70B model that takes ninety seconds to load is a

ninety-second outage from the user's perspective — but the same fundamentals shape steady-state inference throughput, training

step time, and checkpoint behavior. We want someone who understands how all of those parts interact.

• Profile and optimize the end-to-end path for containerized ML/AI workloads: image distribution, runtime startup, weight

loading, inference hot path.

• Design and tune the storage layer that sits between S3-compatible object stores and GPU nodes — prefetchers, caching tiers,

network filesystems and local NVMe layouts.

• Drive measurable wins on time-to-first-token, training step time, and cold-start latency across both internal services and

customer workloads.

• Hands-on experience with at least one distributed or network filesystem and with S3-compatible object storage — including

performance-killing cases (small-object overhead, range-request patterns, eventual consistency, multipart tuning).

• An end-to-end view: comfortable reasoning about NIC, switch, filesystem, cache, container runtime, and GPU as one system rather

than separate silos.

• Experience with container image acceleration (eStargz, SOCI, Nydus) or building a caching layer (FUSE-based, in-kernel, or

sidecar) for ML weights and datasets.

Cash + equity compensation along with various fringe benefits. A team that takes performance seriously, real GPUs to test against,

and a problem space where the wins are visible to every customer.