About the job
Our Edge Inference team compiles Liquid Foundation Models into optimized machine code that runs on resource-constrained devices: phones, laptops, Raspberry Pis, and watches. We are core contributors to llama.cpp and build the infrastructure that makes efficient on-device AI possible. You will work directly with the technical lead on problems that require deep understanding of both ML architectures and hardware constraints. This is high-ownership work where your code ships to production and directly impacts model performance on real devices.
Responsibilities
Implement and optimize inference kernels for CPU, NPU, and GPU architectures across diverse edge hardware
Develop quantization strategies (INT4, INT8, FP8) that maximize compression while preserving model quality under strict memory budgets
Contribute to llama.cpp and other open-source inference frameworks, including new model architectures (audio, vision)
Profile and optimize end-to-end inference pipelines to achieve sub-100ms time-to-first-token on target devices
Collaborate with ML researchers to understand model architectures and identify optimization opportunities specific to Liquid Foundation Models
Qualifications
Minimum
5+ years of experience in systems programming with strong C++ proficiency
Embedded software engineering experience or work on resource-constrained systems
Understanding of ML fundamentals at the linear algebra level (how matrix operations, attention, and quantization work)
Experience with hardware architecture concepts: cache hierarchies, memory bandwidth, SIMD/vectorization
Preferred
Contributions to llama.cpp, ExecuTorch, or similar inference frameworks
Experience with Rust for systems programming
Background in custom accelerator development (TPU, NPU) or work at companies like SambaNova, Cerebras, Groq, or Google/Amazon accelerator teams
Quantitative degree (mathematics, physics, or similar) combined with engineering experience