Software Architect: AI Systems & Co-Design
Software Architect - AI Systems & Hardware Co-DesignWe're working with a well-backed, seed-stage company operating at the frontier of AI and semiconductor design. Our client is solving a genuinely novel infrastructure problem in the AI space. The founding team has deep roots at top-tier hardware companies and a strong research pedigree in AI-aided chip design.This is a ground-floor opportunity - you'd be one of a small number of technical hires shaping the direction of the company.The RoleYou'll sit at the intersection of frontier AI software and specialized silicon - mapping complex model workloads to custom hardware architectures and defining how the two co-evolve. Your work will directly impact the efficiency and scalability of next-generation AI systems.What You'll Be DoingArchitecting how LLM and multimodal workloads map to specialized AI accelerator hardware.Optimizing inference kernels - attention mechanisms (FlashAttention), KV cache management, quantization (INT4), and sparse tensor compression.Defining hardware-software interfaces with a focus on energy efficiency and low-power edge performance.Conducting PPA trade-off analysis across compute density, memory bandwidth, and on-chip SRAM.Contributing to AI-aided design research to accelerate the architectural design loop.What They're Looking For5+ years in AI accelerator software-hardware co-design.Hands-on experience with NVDLA, Tensor Cores, or custom NPUs - real silicon, not simulated.Strong background in LLM inference optimization on GPUs - attention kernels, memory management, quantization.Deep understanding of HBM economics and memory bandwidth constraints in AI infrastructure.Master's or PhD in EE or CS preferred - especially with a research focus on AI-aided hardware design.Familiarity with emerging architectures such as SSMs (Mamba).If you're a systems-level engineer who has operated at the boundary of software and silicon - and you want to work on something foundational rather than incremental - APPLY NOW