Own the serving layer of GoodData Native Inference, deploying and optimizing LLMs on GPUs for production use.
Posted by employer 21 hours ago
First seen on Joblaze 8 hours ago
Last verified on the company career page 8 hours ago
Skills & Technologies
What you'll build
Must have
Nice to have
AI in the day-to-day
We run our AI on open-weight models on our own GPUs, ensuring customer data remains secure.
Requirements
Not disclosed in this posting: compensation, visa sponsorship.
Joblaze summary
In this role, the Sr. Inference Engineer at GoodData.AI is responsible for managing the serving layer of their Native Inference platform, ensuring that AI models operate efficiently in production. Key skills include hands-on experience with modern LLM serving engines and a solid understanding of inference bottlenecks, as well as proficiency in Docker and Kubernetes. This position is ideal for someone with over five years of software engineering experience, particularly in machine learning, who can communicate effectively with both technical and product teams.
Joblaze insights
Quick facts
From the original posting
Our platform turns governed business data into answers, reports and agents that companies put in front of their own customers. To make that AI trustworthy, private and affordable, we run it on our own inference: open-weight models on our own GPUs, so customer data never has to leave our boundary.
You will own the serving layer of GoodData Native Inference. That means the models, the engine and the GPUs behind every AI answer our customers get. You will join the team that builds GoodData AI and become the person the whole company turns to for anything about running models in production.
Deploy, operate and upgrade open-weight LLMs and smaller task models (guardrails, PII detection, classifiers, judges) on our GPUs, in production and across regions.
Make inference fast and cheap: tune batching, KV and prefix caching, quantization and parallelism, and track time to first token, tokens per second per GPU, p95 latency and cost per answer.
Choose the models: put new open-weight models through our evaluation harness on real analytics workloads, and recommend what goes to production.
Grow our inference gateway: routing by intent and complexity across models, a semantic cache, and validation that escalates weak answers.
Adapt models to our use cases with LoRA adapters and fine-tuning, such as generating queries for our semantic layer, root-cause analysis and narrative reports.
5+ years of software engineering, including 2+ years running ML or LLM models in production.
Hands-on production experience with at least one modern LLM serving engine: vLLM, SGLang or TensorRT-LLM.
A working understanding of inference bottlenecks (KV cache, attention, memory bandwidth, batching, parallelism), and the habit of reasoning about latency, throughput and cost in numbers.
Experience with Docker, Kubernetes and GPU instances in a public cloud.
Discipline in benchmarks and evaluation: you do not ship a model you have not measured.
Clear communication in English, with engineers and with product people.
Speculative decoding, multi-LoRA serving, or PEFT fine-tuning.
Contributions to vLLM, SGLang, PyTorch or similar open-source projects.
CUDA or Triton kernels, and the NVIDIA GPU stack (drivers, NCCL).
LLM routing, guardrail models, embeddings and rerankers.
Go or Rust; Terraform and Helm.
Observability for LLM systems (OpenTelemetry, Langfuse, Prometheus, Grafana).
Experience with analytics, semantic layers or text-to-SQL.
Python, Kubernetes on AWS, NVIDIA GPUs, SIE (an open-source inference engine on SGLang), vLLM, PyTorch, gRPC, Langfuse, and our own evaluation harness.
#LI-Hybrid
Standard company text repeated across GoodData.AI's postings is omitted here.