The posting
Description:
Build the systems that turn frontier models into reliable scientific agents capable of tackling complex problems in drug discovery. Own the infrastructure that lets them run experiments, work with scientific data, and carry investigations from question to result.
About Axiom:
Axiom’s mission is to eliminate unexpected drug toxicity and unlock an abundance of ultra safe medicines. We are starting with the liver, where toxicity remains a leading cause of drug failure. Our high-throughput human liver safety platform combines advanced human cell systems with imaging, transcriptomics, mass spectrometry, and drug metabolism measurements. Our AI agents integrate these data with chemical structure, human exposure, and thousands of clinical reference compounds to predict whether a drug will be toxic in humans, why, and how to make it safer.
We work with seven of the world’s top twenty pharmaceutical companies to design safety into molecules from the earliest stages. This applied work feeds directly into our partnerships with leading frontier AI labs, turning real discovery challenges and experimental feedback loops into scientific skill trees and reinforcement learning environments that train increasingly superhuman scientists. We then deploy these frontier models to make cancer therapies safer, unlock treatments for chronic diseases that demand exceptional safety, revive promising drugs abandoned because of toxicity, and much more.
Our ultimate ambition is to build a superhuman AI agent that predicts how a drug will behave across all major human organ systems before the first patient receives it. With insanely accurate safety predictions, we aim to enable rapid, high-throughput testing of novel drugs in humans with zero risk of toxicity. This is the key to unlocking an abundance of ultra safe, effective medicines.
What you will do:
- Build the runtime, tools, and orchestration that enable frontier models to carry out long-horizon scientific investigations, including parallel agents, durable state, and recovery from failures.
- Create fast, secure, sandboxed environments where agents can run Python, analyze data, invoke models, and produce artifacts, with reproducibility and customer isolation built in.
- Develop retrieval, memory, and context compaction systems that connect agents to authoritative data sources and scientific tools across chemistry, biology, and clinical research.
- Make every agent run observable and replayable by capturing model calls, tool calls, state transitions, inputs, and outputs. Build debugging tools to understand why a run succeeded or failed.
- Build rigorous evaluation systems using expert-reviewed reference datasets, production failure cases, model-based judges, regression tests, and automated quality gates.
- Own operational reliability, including retries, timeouts, budget limits, permissions, output verification, and versioning. Build safeguards that catch failures and prevent silent scientific and data errors.
- Work directly with scientists and ML researchers to translate expert judgment into better agent tools, evaluation rubrics, and reliable scientific workflows.
- Bring research agents into the customer-facing Translate product and support reinforcement learning research through agent rollouts, task environments, reward instrumentation, and training data pipelines.
What we look for:
- You’re an exceptional software engineer who has built clean, reliable production systems, with depth in infrastructure, platform engineering, distributed systems, or developer tools.
- You have hands-on experience with LLMs, tool-using agents, or closely related systems. You’ve debugged complex workflows and understand how failures can compound over long-running tasks.
- You take ownership from problem to production. You identify what matters, make progress through ambiguity, and ship systems people rely on.
- You exercise strong engineering judgment. You simplify hard problems, avoid unnecessary abstractions, and build systems that remain understandable as they grow.
- You make decisions with evidence. You define what success looks like and build the evaluations, observability, and feedback loops needed to measure it.
- You care deeply about correctness and reliability. You recognize when an unnoticed failure in an environment, dataset, or tool could undermine a scientific conclusion, and design systems to catch it.
- You’re intellectually curious and comfortable learning across disciplines. You work closely with chemists, biologists, and ML researchers to understand their judgment and translate complex scientific workflows into reliable software.
Technical skills we value:
- Strong Python and backend engineering skills, including asynchronous execution, API development, FastAPI, testing, and production service design.
- Experience with cloud infrastructure and isolated execution environments using tools such as Modal, Docker, AWS, and Terraform.
- Experience building data systems with DuckDB or MotherDuck, Postgres, object storage such as S3, and pipelines for processing and retrieving large scientific datasets.
- Experience building LLM applications and agents, including model APIs, tool use, orchestration, context management, sandboxing, tracing, and evaluation.
- Familiarity with research infrastructure, including experiment tracking, agent rollouts, reward instrumentation, and tools for observing and debugging agent behavior.
- Experience with TypeScript and frontend frameworks such as SvelteKit, Svelte 5, or React is a plus.
