About the company
Our client is a fast-growing chemicals company.
The role
Raydar is recruiting for this role on behalf of our client. Lead research work from early hypothesis to working model, tying gains from offline experiments to results seen in practice. Work across model training, data pipelines and evaluation design, partnering with domain specialists to turn findings into better models. Well suited to a senior researcher who has trained large models hands-on and has turned unclear questions into systems that others rely on.
What you'll do
- Modify and train large models, from additional base-training passes to instruction tuning, selecting data blends and objectives suited to specialized data.
- Build reinforcement learning methods that learn from real-world outcomes despite delayed, noisy and limited feedback.
- Design synthetic-data and knowledge-distillation pipelines with careful provenance tracking, filtering and quality controls.
- Create benchmarks and evaluations that test generalization to unfamiliar problems and connect offline results to practical performance.
- Run and troubleshoot large distributed training jobs, improving throughput, checkpointing and the cost of serving models.
- Study learned models for prediction, search and planning using controlled ablations and strong baselines.
- Work with domain specialists and product engineers to set priorities, analyze failures and turn results into model improvements.
- Help set the research agenda, raise engineering practice, mentor colleagues and share work through papers or open-source releases where suitable.
Requirements
What we're looking for
- Proven experience building and refining large neural networks, with a clear account of what drove results and what did not.
- Mastery of a core technique, whether learning from rewards, scaling language models, generating training data or planning with learned world models.
- Fluent coding skills for research in a Python-based deep learning stack.
- Hands-on experience with distributed training frameworks such as FSDP, DeepSpeed or Megatron.
- Skill in building repeatable tests that catch shortcuts, such as leaked test data, gamed rewards or shifts in input patterns.
- A history of carrying open-ended research questions all the way to production-grade models that others depend on.
- Five to ten years in research or engineering roles, with doctoral study counting toward that experience.
- Ability to learn an unfamiliar technical domain quickly.
Bonus points
- Experience with scientific data, multimodal data or experimental science.
- Familiarity with learning from fixed datasets, judging model confidence, choosing which experiments to run next, or planning when the model is imperfect.
- Skill in making models faster and cheaper to run, such as through weight compression or lower-precision arithmetic.
- A record of papers, public code or shipped systems that show inventive problem-solving and solid delivery.
Benefits
Compensation and benefits
- Base salary: USD 200,000 to 300,000 per year
- Equity
- Health, medical and dental coverage
Location and work model
- San Francisco, CA, United States
- On-site
- Full-time
Skills Required
- Proven experience building and refining large neural networks
- Mastery of a core technique such as reward learning, language-model scaling, training-data generation, or learned-world-model planning
- Fluent Python-based deep learning research coding skills
- Hands-on experience with distributed training frameworks such as FSDP, DeepSpeed, or Megatron
- Experience building repeatable tests that detect data leakage, gamed rewards, and input-pattern shifts
- History of taking open-ended research questions through production-grade models
- Five to ten years of research or engineering experience, with doctoral study counting toward experience
- Ability to quickly learn an unfamiliar technical domain
- Experience with scientific, multimodal, or experimental data
- Familiarity with fixed-dataset learning, model confidence, experiment selection, or imperfect-model planning
- Experience improving model performance and cost through weight compression or lower-precision arithmetic
- Record of papers, public code, or shipped systems demonstrating inventive problem-solving and delivery
What We Do
Raydar is a talent acquisition and business consulting firm that connects world-class and emerging talent with growing organizations. It supports companies through team development, strategic hiring, and customized growth solutions, helping clients recruit roles such as engineers, product managers, executives, legal counsel, and quantitative traders. Raydar focuses on understanding each organization’s needs, culture, and long-term goals to build high-impact teams.
.png){kind=logo}
