Sr. Staff ML Researcher, LLM Algorithmic Optimization

The role

This person will develop new methods for running large language models efficiently on DNN accelerators. They will research and implement algorithmic techniques to improve latency, throughput, memory efficiency, and accuracy.

What you will do

• Research and develop algorithmic and numerical techniques to improve LLM inference on d-Matrix DNN accelerators.
• Design and evaluate methods such as low-precision quantization, model compression, sparsity, pruning, distillation, and low-rank approximation.
• Optimize transformer inference workloads including matrix operations, attention mechanisms, and KV-cache utilization.
• Analyze tradeoffs among model quality, numerical precision, memory footprint, compute efficiency, and system performance.
• Build research prototypes and evaluation frameworks in Python.
• Partner with hardware, compiler, and software teams to map algorithms to accelerator capabilities.
• Stay current with advances in efficient LLM inference, neural-network optimization, numerical methods, and hardware-aware machine learning.

Required qualifications

• Master's degree or PhD in Computer Science, Mathematics, Statistics, Physics, Electrical Engineering, or a related quantitative field
• 5+ years of relevant hands-on experience
• Strong foundation in machine learning and applied mathematics
• Hands-on experience developing and evaluating machine-learning algorithms, models, or research prototypes
• Strong Python programming skills
• Experience with object-oriented code design
• Experience with modern machine-learning frameworks such as PyTorch, JAX, or TensorFlow
• Experience improving the performance, efficiency, or quality of machine-learning models through algorithmic or numerical techniques
• Ability to design rigorous experiments, analyze results, and communicate technical findings clearly

Nice to have

• Experience with transformers, LLM inference, or hardware-aware ML optimization

About the company

d-Matrix has pioneered a memory-centric architecture that brings compute directly into memory. The team works on chips, software, and rack-scale systems to tackle AI infrastructure challenges.