Principal Research Scientist (AI)

TL;DR

Principal Research Scientist (AI): Developing best-in-class realtime multimodal models and the only realtime orchestration platform optimized for thousands of queries per second with an accent on training, new architectures, RL, reward modeling, and scaling. Focus on evaluation, benchmarks, quality measurement, LLM-as-judge, and failure analysis.

Location: Must already have the legal right to work in the United Kingdom, as visa sponsorship is not available for this role. Relocation to the San Francisco Bay Area may be available in the future.

Company

Inworld is a product-oriented research lab of top AI researchers and engineers, developing best-in-class realtime multimodal models and the only realtime orchestration platform optimized for thousands of queries per second.

What you will do

• Hand unclear problems and make them clear.
• Value researchers who say "I don't know yet" and then design the experiment that finds out.
• Treat evaluation as a first-class research product, not a box to check before launch.
• Impact comes before publications though supporting sharing work that moves the field forward.
• Ensure your work is visible.
• Work within a flat structure with fast iterations and minimal process theater.

Requirements

• Experience with foundation models, evaluation, and frontier topics.
• Published research at ICML, ICLR, NeurIPS, EMNLP, ACL, or AAAI.
• PhD in ML/NLP or equivalent practical experience.
• Public work: non-trivial AI side projects, interdisciplinary experiments, or open-source contributions.
• Full-stack research ownership: you frame the question, run the experiments, write the paper, and ship the result.
• Must already have the legal right to work in the United Kingdom, as visa sponsorship is not available for this role.

Culture & Benefits

• Expect unclear problems and the opportunity to make them clear.
• Value researchers who embrace uncertainty and design experiments to find answers.
• Evaluation is treated as a first-class research product.
• Impact is prioritized over publications.
• Flat structure, fast iterations, and minimal process theater.