The posting, in Together AI's own words
archived Sep 9, 2026About the Role
We're looking for a software engineer to build the Kubernetes-native control plane that provisions and runs our GPU inference fleet. You'll design a manifest-driven API where the inference team declares what they need, whether that's a cluster, a model deployment, or a capacity change, and our controllers handle the reconciliation, provider/runtime selection, and lifecycle management underneath, so the inference team never has to know or care which specific serving stack, scheduler, or hardware pool is doing the work. You'll also build the systems that keep the fleet efficient, not just running, including defragmentation and rebalancing logic that consolidates scattered workloads back into contiguous capacity, and scheduling/bin-packing improvements that push GPU utilization up without hurting latency. The core value we're after is decoupling the people building on top of the platform from the operational and runtime complexity underneath, while squeezing more usable capacity out of the same hardware. You'll build the controllers, reconciliation loops, and self-service surface (API/CLI, not tickets) that make that decoupling real, plus the event-driven health, remediation, and utilization systems that keep it running and efficient without a human in the loop. Strong candidates have hands-on experience with Kubernetes controller/CRD patterns, have built or operated a platform API that abstracts multiple backends behind one interface, understand GPU scheduling and capacity efficiency (fragmentation, bin-packing, right-sizing), and think about GPU infrastructure as software to be engineered. A product mindset - you've built internal platforms or APIs consumed by other engineering teams and care about the developer experience of what you ship. You build it, you own it. You are not only responsible for delivering the software but also for operating and supporting it in production.
Responsibilities
Build the provisioning state machine: design and implement the software that models the full lifecycle of a physical host from discovery, inference bring-up to GPU driver/CUDA stack, health validation, and decommission/RMA — as explicit, versioned states and transitions. Build the self-service API: design declarative APIs and a control plane so the inference team can request, scale, and tear down inference clusters with one API call — no ticket, no human in the loop. Automate self-healing: detect degraded or failed nodes, drain them safely, trigger repair or replacement, and reintroduce healthy capacity into the pool automatically. Own reliability of the pipeline: idempotency, retries, rollback, and drift detection so the provisioning system is as dependable as any other production service. Partner with the inference/ML platform team: understand the cluster shapes they need — topology, interconnect, scheduling constraints — and encode them as first-class abstractions in the platform. Engineer it like software: strong typing, automated tests, code review, versioning, and for infrastructure code — this is a product, not a collection of Ansible playbooks.