AI Researcher

The core challenge is turning scientific knowledge and noisy, incomplete laboratory observations into reliable decisions about what to make and how to make it. This hire will help connect model predictions to experimental outcomes so Chemstack AI can reason about chemistry, plan experiments, and improve through feedback.

You will take research ideas through training, evaluation, and deployment for Chemstack AI. The primary spike is building a reliable learning loop from chemistry data and laboratory results: selecting and adapting models, shaping data and post-training methods, evaluating generalization, and connecting offline model improvements to real experimental performance.

What You'll Own

• Build and adapt the model-training stack, including continued pretraining, supervised fine-tuning, data mixtures, molecular representations, and chemistry-specific training objectives.
• Develop reinforcement learning from experimental feedback using reaction outcomes, yields, and process constraints while accounting for delayed feedback, measurement noise, and limited experimental budgets.
• Create synthetic-data, expert-trace, and distillation pipelines with provenance, filtering, and quality controls across literature, reaction records, and laboratory observations.
• Build benchmarks for synthesis planning, procedure generation, and tool use; evaluate generalization to unfamiliar chemistry and connect offline results to laboratory performance.
• Own distributed training runs, diagnose instability, improve throughput and checkpointing, and work with engineers on practical model latency and serving cost.
• Explore learned reaction and process models for outcome prediction, search, and planning using ablations and strong baselines.
• Work directly with chemists and product engineers to prioritize problems, analyze failures, and turn findings into better models.
• Contribute to research direction, engineering standards, mentoring, publications, or open-source work where appropriate.

Technical Requirements

• Strong record of training or post-training large models, with substantive experiment ownership and the ability to explain why approaches worked or failed.
• Deep expertise in at least one of reinforcement learning and reasoning, large-scale language-model training, synthetic data and distillation, or world models and model-based planning.
• Strong research programming in Python and PyTorch or JAX.
• Hands-on experience with distributed training systems such as FSDP, DeepSpeed, or Megatron.
• Ability to design reproducible evaluations and investigate data leakage, reward exploitation, and distribution shift.
• Experience taking an ambiguous research problem through to a dependable model that people use.

Execution and Ownership

• Own work from research hypotheses through training, evaluation, and deployment.
• Use careful experiments, ablations, reproducible evaluation pipelines, and strong baselines to guide investment decisions.
• Move between open-ended research and reliable engineering while diagnosing instability and improving training and deployment systems.
• Collaborate closely with chemists and product engineers to prioritize problems and interpret experimental evidence.
• Communicate technical tradeoffs clearly and change direction when evidence calls for it.

Background

• Typically 5+ years of relevant research or engineering experience, including doctoral research.
• Demonstrated contributions are valued over a particular degree or career path.
• Chemistry experience is welcome, but the role requires the ability to learn the domain.
• Relevant experience may include RL post-training with verifiable rewards, tool-using agents, multi-step reasoning, sparse or delayed feedback, molecular representations, scientific multimodal data, reaction prediction, computational chemistry, experimental science, offline RL, uncertainty estimation, active learning, planning under imperfect learned dynamics, efficient inference, quantization, model compression, or serving optimization.

Location and Visa

• Full-time, in-office role based in San Francisco, CA.
• New hires are expected to work from Mstack's San Francisco office and at least one interview conversation must take place in person.
• Visa transfers or renewals may be supported; candidates requiring a lengthy overseas sponsorship process, such as a six-month wait, are not preferred. Sponsorship needs should be flagged early.

Nice to Have

• Molecular representations, scientific multimodal data, reaction prediction, computational chemistry, or experimental-science experience.
• Offline reinforcement learning, uncertainty estimation, active learning, or planning under imperfect learned dynamics.
• Efficient inference, including quantization, model compression, and serving optimization.
• Publications, open-source contributions, or deployed systems demonstrating original thinking and strong execution.
• A chemistry background.

Who Will Thrive Here This role suits someone who wants to build for science, enjoys working closely with domain experts, and is motivated by experimental evidence. Mstack is looking for people who seek to understand difficult problems deeply, communicate their reasoning clearly, and can contribute across research and engineering as an early member of the AI team.