Member of Technical Staff, Post-Training

• Develop and tune post-training recipes. Design and iterate on post-training stages, datasets, reward signals, and hyperparameters for biological world models. Study how choices in data mixtures, objective design, curriculum, and training schedules affect model behavior.
• Build evaluations that actually matter. Collaborate with the science team to develop and refine evaluation suites for biological reasoning, scientific usefulness, long-context behavior, robustness, and model reliability, and to identify when existing benchmarks stop being informative and should be replaced with better ones.
• Debug model behavior end-to-end. Investigate failure modes in training runs and model outputs, distinguish between signal and noise, and trace problems back to data, optimization, evaluation design, or systems issues.
• Work on preference- and feedback-driven learning. Explore methods such as preference modeling, reward modeling, synthetic feedback, or related post-training approaches that improve how models respond to scientific tasks and constraints.
• Improve data for post-training. Help define, curate, or generate high-quality post-training datasets, including expert-informed data, synthetic data, and task-specific examples grounded in biological workflows.
• Study scaling in post-training. Measure how performance changes with dataset size, recipe complexity, compute budget, and model family. Use those results to guide what we scale next and what new directions are worth exploring.
• Collaborate across research and engineering. Work closely with colleagues in training systems, architecture, and biology-facing research to ensure post-training methods are grounded in the realities of large-scale experimentation and downstream scientific use.

• Strong track record in ML research or engineering, especially in frontier-model training, post-training, alignment, evaluation, data quality, or related areas.
• Proficiency in building production-quality software and research infrastructure, ideally in Python and PyTorch, with comfort debugging large-scale training workflows.
• Ability to design careful experiments, interpret ambiguous results, and separate real effects from artifacts, bugs, or benchmark overfitting.
• Excellent written and verbal communication skills, especially the ability to explain technical findings clearly across research, engineering, and scientific collaborators.
• Curiosity, rigor, and a bias toward iteration: you like improving systems by repeatedly tightening the loop between hypotheses, experiments, and insight.

• Experience with RLHF, RLAIF, preference optimization, reward modeling, rejection sampling, or other post-training methods for large models.
• Experience designing or operating evaluation frameworks for model quality, reliability, safety, or scientific task performance.
• Familiarity with synthetic data generation, annotation workflows, or expert-in-the-loop data collection.
• Background in applied math, systems, computational biology, or another quantitative scientific field.
• Contributions to open-source ML systems, model tooling, or research infrastructure.

• Help build the post-training stack for multimodal biological world models that could materially improve how we detect, understand, and respond to problems in health and biology.
• Work in an environment that combines distributed systems, model architecture, and numerics research with real biological applications.
• Join a collaborative culture that values rigor, creativity, and cross-disciplinary partnership across AI labs, biotechs, hospital systems, and research institutes.
• Competitive compensation, comprehensive benefits, and support for continual learning.

Radical Numerics http://radicalnumerics.ai is an AI research lab building general biological intelligence. Our mission is to master the code of life, and our purpose is to reduce human suffering.

Our team created Evo, and started the field of generative genomics. Our work was featured on the cover of Science https://www.science.org/doi/10.1126/science.ado9336, and presented by our CEO on the main stage of TED2025 https://www.youtube.com/watch?v=EnbfoFUFm2s. Evo was used to create the first AI-generated CRISPR-Cas9, and the first AI whole genome https://www.biorxiv.org/content/10.1101/2025.09.12.675911v1 from scratch. Evo 2 https://www.nature.com/articles/s41586-026-10176-5, featured in Nature, is the largest fully open source AI project across any domain.

In October, we introduced Omnii-1, the world's most advanced multimodal genome language model, and a leap toward general biological intelligence. It unifies 28 biological modalities, spanning DNA, RNA, proteins, epigenomics, and more, in a single model trained on trillions of tokens of biological data In a world first https://x.com/exnx/status/2112580926651904379, Omnii-1 designed the first AI-generated bacteriophages to be administered to a human patient with an antibiotic-resistant infection under compassionate use.

We released this news alongside seven other project areas https://www.radicalnumerics.ai/radical-advances, showing the model’s breadth of applications.

The same generative breakthroughs that enable life-saving cures also lowers the barrier to creating engineered threats and AI-generated bioweapons. We believe these forces are inseparable. Radical Numerics was founded to develop both the power to design and the responsibility to defend.