Founding Engineer, AI Infra

About Goaly

About the Role

Key Responsibilities

• Efficiency & performance: Improve LLM training and inference efficiency through better memory utilization, optimized parallelism, and kernel-level innovations (e.g. FlashAttention, CUDA/Triton).
• Training & RL robustness: Build scalable, stable training and RL pipelines with strong reproducibility, observability, and debuggability.
• Serving & inference optimization: Design and tune high-throughput, low-latency model serving systems, including quantization, caching, and speculative decoding.
• Scalability & infrastructure: Own end-to-end training and inference infrastructure — from data ingestion and checkpointing to multi-GPU and multi-cloud orchestration.
• Production enablement: Work closely with researchers and product engineers to turn new algorithms into reliable, production-ready systems.

Requirements

• 5+ years building or operating ML infrastructure at scale, ideally supporting large language or multimodal models.
• Deep understanding of GPU architecture, distributed training frameworks (PyTorch, DeepSpeed, Megatron, Ray), and parallelism strategies.
• Hands-on experience running inference stacks (vLLM / SGLang, TGI, Triton) and optimizing them via low-level profiling.
• Strong software engineering fundamentals in Python and one of C++/Rust/Go, with clean, reliable code shipped to production.
• Working knowledge of modern data pipelines, feature stores, and vector databases used in production AI systems.
• Comfort automating infrastructure with Kubernetes, Terraform/Pulumi, and observability stacks (Prometheus, Grafana, OpenTelemetry).

Bonus Points

• Experience deploying open-source LLMs (Llama 3, Qwen, DeepSeek) or training custom foundation models.
• Contributions to ML systems tooling (compilers, kernels, inference runtimes) or open-source infrastructure projects.
• Background in reinforcement learning, evaluation harnesses, or alignment tooling that hardens production AI systems.