About the roleApplications will be reviewed continuously, with priority consideration for submissions received by December 7, 2026.You will play a major role in advancing the foundation model and building the agentic systems behind Foundry, the AI platform that informs every drug decision at Pathos, and make them continually improve over time from real-world feedback.
You will own how we train and adapt frontier open-weight models specifically for drug development. You will leverage 200+ petabytes of multimodal, multi-scale oncology data spanning over 8 million patient records linked to longitudinal outcomes, alongside live data streams from clinical trial sites. Supported by a dedicated large-scale GPU cluster and our internal wet lab for experimental validation, your models will directly inform design and execution of clinical trials. You will operate within a lean, high-velocity team where every builder maintains end-to-end ownership over key platform capabilities.
If you want to build the intelligence behind how new medicines are developed, and prove rigorously that it gets better, this is the place to do your career's best work.
What You'll Build- Advance mid- and post-training of our oncology foundation model across SFT and RL, including reward design and verifiers for scientific and clinical tasks whose correctness is hard to check automatically.
- Run and debug training at scale, addressing instability, divergence and reward hacking.
- Build the harness for scientific and clinical workflows: memory, context management, orchestration and tools, plus methods that optimize the harness automatically and evolve it together with the underlying model.
- Design evaluations for open-ended, long-horizon scientific, clinical and operational tasks, calibrate them against expert judgment, investigate where they fail, and use them to decide what ships.
- Build pipelines that generate, filter and curate training and evaluation data, with safeguards for quality, provenance and contamination.
- Turn expert feedback, wet-lab results and real-time signals from clinical trial sites into improvements to models, harness and evaluations, so the platform learns from experience with very high sample efficiency.
- Work directly with oncologists, biologists and the clinical team to translate drug development needs into tractable ML problems, and publish in top ML and scientific venues.
Who You AreWe're looking for researchers and engineers who have gotten models or agents working at scale, care as much about the drugs as about the models, and want their work to reach patients. You take ambiguous problems through experiments to results people can use.
We're hiring across levels, from exceptional early-career talent to experienced researchers.
Must-haves- Deep, hands-on expertise in at least one of the following, which we value equally:
- Mid- or post-training at scale, such as SFT, RL, reward modeling or preference optimization, with direct ownership of training runs and results.
- Evaluation, data and agent harness design, such as evaluation and feedback design for scientific, clinical and operational tasks, synthetic data pipelines, or harness optimization spanning memory, context management, orchestration and tools.
- Rigorous scientific judgment, including when ground truth is limited, delayed or disputed.
- Strong engineering skills in Python, with production-quality code.
- Clear communication of evidence, limitations and uncertainty.
- Willingness to learn biology and drug development, and work closely with domain experts.
Publications, research software and deployed systems are all valid evidence. Publications are welcome but not required.
Bonus- You've built self-improving agentic systems, for example through self-play, synthetic data generation, self-training on past trajectories or test-time learning, or systems that automate ML research.
- You've built AI systems for real-world scientific workflows, such as agents that generate hypotheses and design, run and analyze experiments using lab or clinical data.
- You've built RL environments or task suites for training agents on realistic, long-horizon tasks.
- You've modeled multimodal biomedical data, such as genomics, transcriptomics, pathology and radiology images, clinical notes and EHRs, ideally combining several modalities in one model.
- You've worked on uncertainty estimation, calibration or causal inference, especially where model outputs inform high-stakes decisions.
- You've built or scaled infrastructure for efficient training and inference, such as distributed training, custom kernels (CUDA or Triton) or high-throughput serving.
- You've shipped ML systems that real users rely on, then monitored them in production and improved them from user feedback.
- You have a background in oncology, drug development or clinical research, or a track record of working closely with clinicians and biologists.
- You've worked with patient-level or clinical trial data under privacy and regulatory constraints, such as HIPAA.
None of these are required. We value depth in one area over breadth across all of them.
LocationThe preferred location for this position is in our New York City HQ, though exceptional remote candidates will also be considered.
CompensationWe are hiring across levels. The anticipated salary range is $220,000 to $350,000 plus equity, and we have flexibility for exceptional candidates.