> Markdown version of [/jobs/ext/4109966-staff-engineer-software-distribution-onprem](https://www.wearedevelopers.com/jobs/ext/4109966-staff-engineer-software-distribution-onprem). Every page supports `.md` or `Accept: text/markdown`. Links point to the HTML versions so they work for humans too. Agent guide: [/agents.md](https://www.wearedevelopers.com/agents.md). --- # Staff Engineer, Software Distribution onPrem - **Company:** Qevlar AI - **Location:** Paris, France (Remote available) - **Contract:** Permanent contract - **Skills:** Amazon S3, Software as a Service, Domain Name System (DNS), Python (Programming Language), PostgreSQL, OpenShift, Ansible, Software Engineering, YAML, Graphics Processing Unit (GPU), Software Distribution, Large Language Models, Kubernetes, TensorRT, Nim (Programming Language), Terraform, Oracle Cloud Infrastructure - **Published:** October 20, 2026 - **Apply:** https://www.indeed.com/viewjob?jk=e07d08299d7a56c9 ## About the Role * Shipped software that someone else installs. This is what we screen hardest on. Operating your own cluster is a genuinely different discipline from handing an artifact to a stranger who runs it on infrastructure you will never see. * Deep Helm expertise - subcharts, conditional templating, values design - plus OCI registries and real exposure to Kubernetes beyond a single distribution: OpenShift and its SCCs, vanilla, RKE2, k0s, whatever you had to survive. * Worked in air-gapped or heavily restricted environments: image mirroring, offline bundles, egress proxies, custom CA bundles, no external DNS. * Replaced managed cloud services with self-hosted equivalents - Postgres operators such as CloudNativePG, MinIO or another S3-compatible store, Vault or External Secrets - and lived with the operational consequences. * Practical software supply chain security: image signing with cosign or equivalent, SBOM generation, distroless or minimal base images, CVE triage. * Strong infrastructure-as-code and CI (Terraform, Ansible, pipeline matrices). Here the test harness is the product. * Hands-on experience serving models on GPU - vLLM, NVIDIA NIM or TensorRT-LLM - including VRAM and concurrency sizing and the NVIDIA GPU Operator / Container Toolkit. * Strong Go and/or Python. You will change application code, not only YAML. * The judgement and the appetite to say no. Staff here means architectural authority and the willingness to use it. * Fluent in French and English. ## Description * Decide what we do not port. Your first deliverable is the on-prem reference architecture and the explicit list of dependencies, services and features that will not cross the boundary. You will hold that line against the rest of engineering, with our backing. * Ship the bundle. A Helm chart and an air-gapped tarball that install on vanilla Kubernetes and on OpenShift, with no internet access, in under four hours - by someone who has never seen our code and has only your documentation. * Replace the cloud underneath us. Self-hosted Postgres, S3-compatible object storage, secret management, identity federation - behind abstractions clean enough that we never fork the product or let on-prem drift away from SaaS. * Package and size the inference layer. Our platform runs agentic investigations, which means sustained LLM throughput. You own the serving stack in the bundle, the VRAM and concurrency sizing matrix customers use to provision their own GPUs, and the distribution and rollback of model weights across an air gap. * Build the test harness that makes releases boring. Ephemeral customer-shaped environments, install and upgrade smoke tests across Kubernetes distributions, identity providers and network modes - until validating a release candidate takes a day, not a week. * Ship clean. Signed images, SBOMs, minimal base images, no critical or high CVEs at release. Our buyers are security teams; they will scan everything we hand them on day one. * Impose the upgrade contract on the whole org. Tested N-3 upgrade paths, backward-compatible migrations, no feature merged without a feature flag, a health check and a support-bundle contribution. Customers will jump from a year-old release straight to the latest, and it has to work. * Keep it boringly simple. A product that ships self-hosted cannot grow a new managed dependency every quarter. You get explicit veto power, and you will need it.