Machine Learning Architect

About Neurophos

We are developing an ultra-high-performance, energy-efficient photonic AI inference system. We’re transforming AI computation with the first-ever metamaterial-based optical processing unit (OPU).

As AI adoption accelerates, data centers face significant power and scalability challenges. Traditional solutions are struggling to keep up, leading to rapidly rising energy consumption and costs. We’re solving both problems with an OPU that integrates over one million micron-scale optical processing components on a single chip. This architecture will deliver up to 100 times the energy efficiency of existing solutions while significantly improving large-scale AI inference performance.

We’ve assembled a world‑class team of industry veterans and recently raised a $110M Series A led by Gates Frontier. Participants include M12 (Microsoft’s Venture Fund), Carbon Direct Capital, Aramco Ventures, Bosch Ventures, Tectonic Ventures, Space Capital, and others. We have also been recognized on the EE Times Silicon 100 list for several consecutive years.

Join us and shape the future of optical computing!

Austin, TX or San Francisco, CA. Full-time onsite position.

We are seeking an experienced machine learning architect to lead the porting and optimization of large language models (LLMs), diffusion models, and other ML applications to our revolutionary optical inference engines. This role is critical to demonstrating the full potential of our metamaterial‑based optical processing units (OPUs) by adapting state‑of‑the‑art AI models to leverage our ultra‑high‑throughput, low‑precision compute architecture. The ideal candidate will bridge the gap between cutting‑edge ML research and novel hardware capabilities, ensuring customers can seamlessly deploy their AI workloads on Neurophos hardware.

• Lead the porting of LLM applications, diffusion models, and visual ML applications to Neurophos optical inference engines
• Adapt models from diverse sources, including Git Hub, Hugging Face, other open‑source repositories, and customer private models
• Work with models in various formats, including PyTorch, Triton, JAX, and emerging frameworks
• Develop and implement quantization strategies to migrate models from higher precision formats (FP8, INT8, and above) to our optimized 4‑bit precision (FP4/INT4) for weights and activations
• Design and execute re‑quantization, retraining, and other model adaptation techniques to minimize accuracy loss during precision reduction
• Create or integrate third‑party tools and workflows for efficient model porting and optimization
• Optimize GEMM operations for high‑throughput execution
• Develop benchmarking methodologies to measure and validate model quality post‑porting, including perplexity metrics and other quality indicators
• Collaborate with hardware and software teams to co‑optimize model architectures for optical compute characteristics
• Publish research papers on novel optimization techniques and methodologies (with appropriate IP protection)

• MS or PhD in Computer Science, Data Science, Machine Learning, Mathematics, or related field
• 7+ years of experience in machine learning engineering with at least 3 years focused on model optimization and deployment
• Deep expertise in neural network quantization techniques, including post‑training quantization (PTQ) and quantization‑aware training (QAT)
• Strong proficiency in PyTorch and familiarity with other ML frameworks (JAX, Triton, Tensor Flow)
• Hands‑on experience with transformer architectures, LLMs, and diffusion models
• Experience with low‑precision inference optimization (INT8, FP8, or lower)
• Strong understanding of GEMM operations and linear algebra optimizations for deep learning
• Experience with model evaluation metrics, including perplexity, accuracy, and benchmark suites
• Track record of successfully deploying ML models on specialized hardware accelerators
• Excellent communication skills with the ability to collaborate across hardware and software teams

• Experience with sub‑8‑bit quantization (INT4, FP4) and mixed‑precision inference
• Familiarity with Hugging Face Transformers library…