About Belvedir
Belvedir's mission is to make the creation and deployment of AI models accessible to everyone. To achieve this we build autonomous systems for cleaning raw data, creating benchmarks, asynchronous RL, test-time training, inference optimization, and much more.
We work with startups, enterprises, and some of the world's leading applied AI companies. We have raised a seed round from an amazing group of notable VCs and angels.
About the role
- Build and operate the training and serving stack for custom models on Kubernetes: distributed training, batching, caching, routing, and autoscaling.
- Drive down cost per token with quantization, speculative decoding, kernel work, and smarter scheduling.
- Profile end to end and fix the bottleneck, wherever it is: GPU, CPU, network, or the model itself.
- Set and hit latency and reliability targets for customer deployments.
- Partner with training on model architectures and formats that are cheap to serve.
What we are looking for
- Experience training or serving LLMs at scale with PyTorch, vLLM, SGLang, or a system you built.
- Strong systems fundamentals: concurrency, memory, networking, and performance measurement.
- Working knowledge of GPU programming (CUDA or Triton) and transformer training and inference internals.
- Fluency in Python plus a systems language such as C++, Rust, or Go, and hands-on experience running GPU workloads on Kubernetes.
- Comfort owning production systems and being on the hook when they break.
- Contributions to an open-source inference engine or kernel library.
- Experience with multi-tenant serving, LoRA serving, or model multiplexing.
- Experience at an early-stage startup or as a founder.
Benefits and logistics
- Competitive salary and equity
- Health, dental, and vision coverage
- Daily lunch and dinner at the office
- Visa sponsorship
How we hire
- Intro call with a founder
- Technical take home
- Paid work trial with the team