Make science programmable.
Dynamical is building the autonomous R&D system for companies developing complex physical products, where scientific agents turn engineering objectives into temporary self-driving laboratories that discover and qualify the materials and processes required to build them.
Can a model choose the experiment that most improves its understanding of the physical world? Can it use that result to strengthen its simulators and verifiers, change the downstream decision, and carry what it learned beyond the conditions it has already seen?
The network connects scientific agents, simulators, instruments, verifiers, and scientists across laboratories. Each campaign turns uncertainty into experiments and carries the resulting evidence into the next decision.
Open roles
In person in New York City.
Founding Research Scientist, Materials and AI
You define how an AI system learns from physical reality. You turn experimental outcomes, uncertainty, and expert judgment into scientific objectives and training signal, and help choose the questions our models and laboratory pursue. You help set the scientific direction of the company alongside the founder.
Science
Founding Research Scientist, Materials and AI
You define how an AI system learns from physical reality. You turn experimental outcomes, uncertainty, and expert judgment into scientific objectives and training signal, and help choose the questions our models and laboratory pursue. You help set the scientific direction of the company alongside the founder.
What you'll do
- Translate how materials scientists reason over experimental data into the rubrics and evaluations agents are measured against, across synthesis, characterization, and qualification workflows.
- Decide what is worth measuring, what counts as proof, and where today's models fail when they touch physical evidence.
- Audit how agents reason over evidence, find where they cut corners, and turn that into the standard of proof the whole system is held to.
- Trace delayed physical outcomes, including failures and nonformations, back to the decisions that caused them, and define what good judgment looks like across a campaign.
- Set the scientific agenda and the research bar.
The hard problems you'll work on
- How should a scientific agent choose the experiment that will most reduce uncertainty when information gain, cost, risk, and qualification value compete?
- How do experimental outcomes, failures, and expert judgment become scientific objectives and training signal without flattening uncertainty or rewarding rubric-shaped behavior?
What we look for
- A PhD or equivalent depth in materials, chemistry, physics, scientific ML, or a related field, with hands-on lab or simulation experience.
- Worked at the boundary of a science domain and machine learning, and can tell a convenient result from a trustworthy one.
- Fluent in how working scientists make calls across a campaign, and able to turn that judgment into something a system can use.
- Want to own a scientific direction from the start.
Founding Platform Engineer
You build the system connecting models, simulators, instruments, experiments, and scientists. You make each physical run legible to the whole learning loop, so its result can update the policy, world model, verifier, and next decision. You help define the interface through which physical R&D becomes programmable.
Platform
Founding Platform Engineer
You build the system connecting models, simulators, instruments, experiments, and scientists. You make each physical run legible to the whole learning loop, so its result can update the policy, world model, verifier, and next decision. You help define the interface through which physical R&D becomes programmable.
What you'll do
- Carry the full loop from ingesting historical experimental data, through compiling it into machine-readable evidence and running evals and environments, to feeding results back into training and the next decision.
- Build the product where scientists and agents plan, run, inspect, and decide together, with traces, evidence packets, eval runs, review loops, and direct-manipulation interfaces over the evidence.
- Own the data models and provenance, the sample lineage, verifier outcomes, and artifacts, so every experiment leaves a record the system can learn from.
- Sit with the people running campaigns, find where their attention is lost, and build for that gap.
- Own product and platform engineering end to end, from data model and APIs to interface and deployment, and shape it with the founder.
The hard problems you'll work on
- What is the right system of record for a physical experiment, one that connects intent, execution, evidence, uncertainty, and the decisions it changes?
- What interface lets scientists and agents plan, inspect, correct, and learn from the same evidence without hiding provenance or consequential judgment?
What we look for
- Built serious software in ambiguous domains like research platforms, data systems, ML infrastructure, developer tools, or scientific software where correctness and usability both mattered.
- Owned a system end to end, from data model and APIs to interface, deployment, and observability.
- Want to build a product and platform from scratch and own it.
- A bias toward simple systems other people can build on.
Founding Research Engineer, Evals and Environments
You build the environments and training loops in which scientific agents learn what to do next. You measure whether a model can gather the right evidence, choose informative experiments, revise its scientific state, and carry what it learns beyond the conditions that trained it.
Evals
Founding Research Engineer, Evals and Environments
You build the environments and training loops in which scientific agents learn what to do next. You measure whether a model can gather the right evidence, choose informative experiments, revise its scientific state, and carry what it learns beyond the conditions that trained it.
What you'll do
- Own the evals end to end, the harnesses, baselines, metrics, and the offline loop that qualifies an agent's judgment before it touches a live instrument.
- Build long-horizon environments where agents inspect physical evidence, call tools and models, make decisions, and get scored against what reality showed.
- Design the state, actions, rewards, verifiers, and visibility rules that turn a messy scientific workflow into something an agent can be trained and evaluated on.
- Turn dead ends, simulator misses, and expert corrections into harder tasks, hard negatives, and training data, and decide what the environment should teach next.
- Work directly with the founder, who has lived this problem, and own the eval and environment surface as the team grows.
The hard problems you'll work on
- How do you build environments that reward agents for gathering the right evidence and revising their scientific state, not just arriving at the right answer?
- What evaluation shows that an agent can carry what it learned beyond its training conditions without leaking the hidden answer from the historical record?
What we look for
- Built RL environments, agent eval harnesses, tool-use systems, simulators, or benchmark suites that had to survive messy data.
- Trained or post-trained LLMs with RL, or designed reward and verifier contracts with human or model trainers.
- Comfortable taking an ambiguous surface and turning it into clean abstractions.
- Want to own evaluation and environment design as a whole.
Reach out.
You will help define the research agenda, the system, and the kind of scientific organization we become.
Email us with the role you are closest to and one or two things you have built. A paper, repository, dataset, product, or failed experiment you learned from tells us more than a resume.