Research Engineer

Cobalt; San Francisco (CA); On-site; USD 120,000 - 160,000; Full time

Benefits offered by this job

Flexible hours

Remote-first culture

Competitive compensation package

Job summary

Cobalt is seeking a Research Engineer focused on post-training optimization and innovative reasoning techniques. This role offers flexibility, allowing you to work remotely or in person, contributing to advanced AI systems and projects.

Ideal candidates should possess strong ML fundamentals and have prior experience in research engineering or production ML systems. Enjoy high-impact work with competitive compensation as you collaborate with talented professionals in a remote-first environment.

Qualifications

• Comfortable training and fine‑tuning LLMs end‑to‑end.
• Real exposure to post‑training methods like SFT and RL fine‑tuning.
• Proven history of shipping research or production ML systems.

Responsibilities

• Conduct research in post‑training optimization and reasoning techniques.
• Develop innovative algorithms for advanced AI systems.
• Collaborate with teams on R&D projects optimizing AI performance.

Skills

Strong ML engineering fundamentals

Post‑training methods experience

Track record in research

Ability to work independently

Interest in applied work with real-world consequences

Education

PhD or Master's in relevant field

Tools

PyTorch

Hugging Face

vLLM

deepspeed/FSDP

Job description

Role Description

This is a Research Engineer role focused on post‑training and reasoning. Full‑time or part‑time; we're open to candidates currently pursuing a PhD or Master's or working in industry. Your core responsibilities will include conducting research in post‑training optimization and reasoning techniques, developing innovative algorithms, and collaborating with cross‑functional teams to apply findings to advanced AI systems. The role also involves analyzing complex datasets, enhancing AI models, and contributing to cutting‑edge R&D projects aimed at optimizing AI performance and interpretability.

What we're looking for:

• Strong ML engineering fundamentals. Comfortable training and fine‑tuning LLMs end‑to‑end (PyTorch, HF, vLLM, deepspeed/FSDP, or similar)
• Real exposure to post‑training methods (SFT, preference optimization, RL fine‑tuning), not just having read the papers
• A track record of shipping research or research‑grade engineering: publications, strong open‑source contributions, or production ML systems at a lab/frontier company
• Comfortable working with a part‑time research lead. You can take a direction and run, surface tradeoffs early, and don't need someone in the room every day
• Excited by applied work in a domain with real‑world consequences (you don't have to come from healthcare; you do have to care about it)

The work spans the full post‑training stack as applied to expert reasoning:

• Designing and running SFT, DPO, and RL (GRPO/PPO and successors) experiments on reasoning traces from our expert network
• Building benchmarks and evals that meaningfully measure clinical and adjudication reasoning — not just final‑answer accuracy, but the reasoning path
• Turning raw expert outputs into high‑quality training datasets: schema design, quality controls, scaling pipelines
• Working directly with customers (frontier labs, healthcare AI companies) on bespoke data and eval engagements
• Publishing where it makes sense

Why Cobalt?

• High‑Impact Work: Your designs will be used by leading institutional investors and the world's most influential organisations
• Innovation at the Frontier: Work at the intersection of artificial intelligence, institutional finance, and cutting‑edge technology
• Remote‑First Culture: Work from anywhere with flexible hours and a globally distributed team
• Growth Opportunity: Shape design culture and build a world‑class design team as we scale
• Competitive Compensation: Industry‑leading salary, equity, and benefits package

Location and work arrangement

• Remote or in‑person. While we're headquartered in San Francisco, we welcome talented data scientists from around the world. Team collaboration occurs across time zones with overlapping core hours