Machine Learning Researcher

About the role

Inference.net trains and hosts specialized language models for companies who want frontier-quality AI at a fraction of the cost. The models we train match GPT-5 accuracy but are smaller, faster, and up to 90% cheaper. Our platform handles everything end-to-end: distillation, training, evaluation, and planet-scale hosting.

We are a well-funded ten-person team of engineers who work in-person in downtown San Francisco on difficult, high-impact engineering problems. Everyone on the team has been writing code for over 10 years, and has founded and run their own software companies. We are high-agency, adaptable, and collaborative. We value creativity alongside technical prowess and humility. We work hard, and deeply enjoy the work that we do. Most of us are in the office 4 days a week in SF; hybrid works for Bay Area candidates.

Key Responsibilities

• Research and experiment with new model architectures to improve quality, efficiency, or capability
• Explore methods to decrease inference latency and improve serving efficiency
• Run experiments with new learning methods, including novel approaches to SFT, RLHF, DPO, and other post-training techniques
• Develop and improve our distillation pipeline for training high-quality models from frontier teachers
• Train models for clients and run evaluations to validate research findings in production settings
• Create robust benchmarks and evaluation frameworks that ensure custom models match or exceed frontier performance
• Stay current with ML research and identify techniques that can improve our platform
• Collaborate with applied engineers to bring successful research into production systems
• Document findings and share knowledge with the team

Requirements

• 3+ years of experience training AI models using PyTorch
• Deep understanding of transformer architectures, attention mechanisms, and model internals
• Hands-on experience with post-training LLMs using SFT, RLHF, DPO, or other alignment techniques
• Experience with LLM-specific training frameworks (e.g., Hugging Face Transformers, DeepSpeed, Megatron, TRL, or similar)
• Strong experimental methodology, including ability to design, run, and analyze rigorous experiments
• Track record of implementing ideas from recent ML papers
• Experience training on NVIDIA GPUs at scale
• Strong foundation in ML fundamentals: optimization, loss functions, regularization, generalization
• Nice-to-have: Publications in ML venues, Experience with model distillation or knowledge transfer, Experience with LLM speed optimization techniques, Familiarity with vision encoders, multimodal models, or other modalities, Experience with distributed training and infrastructure at scale, Contributions to open-source ML projects