We use AI to discover new pesticides, because the ones we have now are failing — pests keep evolving, and the chemicals kill too many other things (like humans). Our models replace slow, expensive lab work, letting us discover new molecules 100× faster. Our first pesticide targets a Spodoptera with a novel mode of action. We’re a small team of young engineers from Caltech & Wolfram Research, with experience in drug discovery ML models, rebuilding agrochemical R&D around AI.

We’re building an autonomous engine for scientific discovery—an AI system that designs, plans, and learns how to learn. This role is for exceptional people who want to invent a better way to optimize products and generalize models over complex search spaces using reinforcement learning, Bayesian methods, and active experimentation.

You’ll be part of the core research team. No side projects. No throwaway intern work. You’ll be expected to think deeply, write clean code, and help us push the theoretical and practical boundaries of intelligent discovery systems.

What You’ll Work On

• Designing and implementing the core decision-making framework.  
• Running experiments with internal ML models.
• Evaluating performance, generalization, and uncertainty of models.
• Coding mock environments for policy learning.

Qualifications

• Strong mathematical foundation, with emphasis on probability theory, information theory, and optimization.
• Understanding of Bayesian reasoning and decision theory (e.g. expected utility, belief state modeling, POMDPs)
• Experience with learning under uncertainty and active learning methods (e.g. Bayesian optimization) 
• Experience with planning and control algorithms (e.g. Monte Carlo Tree Search, multi-armed bandits such as UCB and Thompson Sampling, meta-RL)
• Knowledge of modeling and generalization techniques, including surrogate modeling, inductive biases, and generalization in high-dimensional search spaces
• Solid practical ML engineering and software engineering fundamentals (Python, PyTorch/JAX/TensorFlow, NumPy/SciPy)
• Ability to mathematically formalize vague problems

Bonus:

• Experience with BoTorch, GPyTorch, Pyro, Stan, or JAX.
• Experience with deep learning methods (e.g. VAE, transformers)
• Prior research or open-source contributions in probabilistic inference, RL, or decision-making.
• Strong performance in math/programming competitions or coursework (e.g. Putnam, IOI, ICPC).

We use Python + PyTorch to build AI models that predict how well molecules will do at killing a pest and nothing else. We use these models to search through billions of possible compounds. Model accuracy and data quality are problems we're always iterating on. Another hard part is choosing which molecules to test in physical reality for data feedback — we're using ideas from information theory and reinforcement learning to pick the highest utility tests, trading off cost and time. On the bio side, we work on designing effective assays that are conducive to large scale data collection for ML, as well as optimizing the process of in-vivo testing on our target pest. Minimizing time between test results is a big priority. Long term, we’re building a general system that can design any molecule for a given task using AI.