Software Engineer, Inference - Multi Modal
San Francisco • FullTime
Posted 1y ago
Remote Work Policy
On-site
Employment Type
FullTime
Categories
Applied AI Engineer
About the job
OpenAI's Inference team is expanding into multimodal inference, building the infrastructure needed to serve models that handle image, audio, and other non-text modalities. These workloads are inherently more heterogeneous and experimental, involving diverse model sizes and interactions, more complex input/output formats, and tighter coordination with product and research. We are looking for a software engineer to help us serve OpenAI’s multimodal models at scale. You’ll be part of a small team responsible for building reliable, high-performance infrastructure for serving real-time audio, image, and other MM workloads in production. This work is inherently cross-functional: you’ll collaborate directly with researchers training these models and with product teams defining new modalities of interaction. You'll build and optimize the systems that let users generate speech, understand images, and interact with models in ways far beyond text.
Responsibilities
- Design and implement inference infrastructure for large-scale multimodal models.
- Optimize systems for high-throughput, low-latency delivery of image and audio inputs and outputs.
- Enable experimental research workflows to transition into reliable production services.
- Collaborate closely with researchers, infra teams, and product engineers to deploy state-of-the-art capabilities.
- Contribute to system-level improvements including GPU utilization, tensor parallelism, and hardware abstraction layers.
Requirements
- Experience building and scaling inference systems for LLMs or multimodal models.
- Experience with GPU-based ML workloads and understanding the performance dynamics of large models, especially with complex data like images or audio.
- Comfort dealing with systems that span networking, distributed compute, and high-throughput data handling.
- Familiarity with inference tooling like vLLM, TensorRT-LLM, or custom model parallel systems.
- Ability to own problems end-to-end and operate in ambiguous, fast-moving spaces.
- Experience working with image generation or audio synthesis models in production (nice to have).
- Exposure to distributed ML training or system-efficient model design (nice to have).