Senior AI Scientist

This is a fully remote position, open to applicants in United States.

📋 Description

• Take charge of the medical modeling roadmap at Atria: determine which model families and training methods are most likely to address our challenges effectively, and develop the models that demonstrate success. The goal is to "move the clinical needle," not merely to "rank on a leaderboard."

• Explore the open-source ecosystem (including general-purpose and clinical/biomedical foundation models, vision and multimodal backbones, and specialized medical models) and provide well-founded recommendations on what to fine-tune, distill, or create.

• Innovate and validate new architectures where existing solutions are inadequate, particularly in the context of longitudinal multi-modal data and the foundation-model approach.

• Oversee fine-tuning and post-training activities: SFT, LoRA/QLoRA and other PEFT techniques, DPO and related preference-tuning methods, RLHF/RLAIF as applicable, ongoing pre-training, and distillation.

• Assemble high-quality training datasets from Atria's clinical data: sampling, labeling strategies, deduplication, contamination checks, train/test hygiene, and PHI-safe handling at all stages.

• Create and implement thorough evaluation processes: held-out clinical benchmarks, comparisons with cutting-edge closed-source baselines, calibration analyses, subgroup performance assessments, and ablation studies to determine which design decisions were effective.

• Manage training and evaluation on suitable infrastructure (single-node and multi-GPU), ensuring proper experiment tracking, reproducible configurations, and logging that enables future reflection on past experiments.

• Remain updated on the latest literature: training methodologies, fine-tuning techniques, medical foundation models, and multimodal architectures. Distill what is pertinent to Atria instead of treating every new arXiv publication as a potential threat.

• Collaborate with clinicians to define what constitutes "good" performance for each model, and partner with researchers for joint training and evaluation efforts.

⛳️ Requirements

• A preference for delivering models rather than writing papers about them. You would rather have a functional v1 ready for clinicians next month than a polished methodology released next year.

• A resourceful attitude. You can quickly learn an unfamiliar fine-tuning technique, training framework, or clinical concept and have a credible experiment prepared within a couple of days.

• A strong commitment to continuous improvement. You actively review others' code, papers, training logs, and post-mortems, treating errors as valuable learning opportunities.

• A graduate degree (PhD preferred, Master's with a robust research background) in computer science, machine learning, computational biology, biomedical informatics, or a related field — or a solid open-source track record in contemporary training and fine-tuning.

• At least 4 years of hands-on experience in training and fine-tuning modern deep learning models, with a record of shipped or published models that you personally trained.

• Comprehensive, up-to-date knowledge of modern fine-tuning and post-training techniques: SFT, PEFT (LoRA, QLoRA, adapters), preference tuning (DPO and successors), distillation, and ongoing pre-training.

• Strong familiarity with the open-source model ecosystem: awareness of which models are state-of-the-art, which are overrated, and what is worth fine-tuning for specific problems.

• Proficiency in Python and PyTorch, with practical experience in the Hugging Face ecosystem (transformers, datasets, PEFT, TRL, accelerate) or equivalent training frameworks.

• Practical experience with training infrastructure: distributed training, mixed precision, efficient data loading, and experiment tracking (W&B, MLflow, or similar tools).

• A disciplined approach to evaluation and ablation: you consider benchmarking, calibration, and "what would this have looked like without that change?" as integral to the modeling process.

• A genuine interest in healthcare and the responsibility associated with developing models that impact patient care.

🏝️ Benefits

• Exceptional health and wellness benefits, fully covered by Atria, effective from the date of hire.

• OneMedical membership for employees and their dependents, providing access to 24/7 virtual care.

• Support for fertility and family planning.

• Company-sponsored preventive health screenings through partner hospitals (e.g., calcium score).

• Fitness perks, including Wellhub +.

• 401k contributions with a 4% match starting after 6 months.

• Flexible time off policy.

• Continuing medical education (CME) and CEU support for professional licensure.