About the job
Join Normal Computing for Incredible Opportunities!
At Normal Computing, we are passionate about building the software and hardware that drive technological innovation. Our efforts support the semiconductor industry, critical AI infrastructure, and the extensive systems that empower our world. With teams spread across New York, San Francisco, Copenhagen, Seoul, and London, we collaborate as one unified team.
Your Role in Our Mission:
We are seeking a talented AI Research Engineer to pioneer advancements in agentic LLMs and reinforcement learning for our innovative code generation tool, Nectar. In this role, you will design and conduct experiments, develop agents, curate datasets from intricate technical documents, and establish rigorous evaluation methods. You will write production-level research code and work closely with engineering teams to deliver enhancements to our clients. While leadership experience is not a prerequisite, your research and development impact will be significant.
Key Responsibilities:
Design and implement multi-agent and reinforcement learning strategies for agentic code generation and tool usage.
Develop research prototypes that integrate with Nectar; collaborate to transition successful prototypes into production.
Create comprehensive evaluation frameworks, including task specifications, pass/fail criteria, coverage metrics, and dashboards for cost and latency.
Gather and curate datasets from a variety of sources, including PDFs, logs, and tables; generate synthetic data where necessary, and maintain data documentation and licensing.
Conduct disciplined analyses of experiments through systematic ablations; document findings and decisions.
Keep abreast of recent developments in LLM agents, reinforcement learning (both offline and online, including RLHF/RLAIF), constrained decoding, and program synthesis.
Qualifications:
PhD in Computer Science, Artificial Intelligence, or Machine Learning (or equivalent research experience), with publications preferably in multi-agent reinforcement learning, agentic AI, or reinforcement learning for language and code.
Proficient in Python and machine learning frameworks, with a preference for PyTorch; experience with JAX or Hugging Face is a plus.
Proven track record of translating research into functional systems; a strong emphasis on reproducibility (tests, seeds, configurations, logging).
Experience in designing evaluation harnesses and success metrics for sequential and agentic tasks.
Adept at data acquisition and curation from various document types; possess strong instincts regarding data quality and licensing.
Bonus Qualifications:
Research experience in program synthesis, code generation, constrained decoding, or execution-based rewards.
