Senior AI Scientist

The Atria Health Institute is a membership-based primary and specialty health care practice with a focus on prevention and longevity. We bring together a multidisciplinary team of renowned physicians to provide proactive, preventive, and precision-based care for Atria members and their families. All care, including primary care, advanced screening and diagnostics, urgent care, specialty care, 24/7 home visits, and imaging is included in members’ annual fee.

We have state-of-the-art facilities with imaging and diagnostics on-site in New York City, Palm Beach, and Los Angeles, with home services hubs in Miami, Greenwich/Westchester, and the Hamptons in summer. Additional locations are forthcoming in the Bay Area, Miami, and in downtown Manhattan. At each of our Institutes, we deliver personalized preventive medicine grounded in up-to-the-minute science — all in serene environments designed for trust and comfort.

Role Overview

As a Senior AI Scientist at Atria, you will own the development of the medical models at the heart of Atria's clinical AI agenda: domain-specific models for the clinical problems that matter most to our practice today, and foundation models for preventive medicine that learn the underlying patterns of how health evolves over years and decades, surfacing risk before symptoms appear and personalizing prevention to the individual rather than the population average. You'll work on a data estate that very few organizations in the world have access to: whole-genome sequencing, advanced imaging including whole-body MRI, comprehensive longitudinal labs, continuous wearable signals, and family-linked records spanning multiple generations. Your core job is to take the best of what the open-source ecosystem has to offer and turn it into models that materially outperform off-the-shelf options on Atria's problems.

Where the problem demands it, you'll design fundamentally new architectures.

What you'll do

• Own the medical modeling roadmap at Atria: identify the model families and training approaches most likely to win on our problems, and build the models that prove it. The bar is "moves the clinical needle", not "places on a leaderboard".
• Survey the open-source landscape (general-purpose and clinical/biomedical foundation models, vision and

multimodal backbones, specialized medical models) and make defensible recommendations on what to fine-tune, distill, or build on.

• Design and validate novel architectures where existing approaches fall short, particularly for longitudinal multi-modal data and the foundation-model direction.
• Lead fine-tuning and post-training work: SFT, LoRA/QLoRA and other PEFT methods, DPO and related preference-tuning approaches, RLHF/RLAIF where appropriate, continued pre-training, and distillation.
• Curate high-quality training datasets from Atria's clinical data: sampling, labeling strategy, deduplication, contamination checks, train/test hygiene, and PHI-safe handling throughout.
• Design and run rigorous evaluation: held-out clinical benchmarks, comparison against frontier closed-source baselines, calibration analysis, subgroup performance, and the ablations that tell you which design choices actually earned their keep.
• Run training and evaluation on appropriate infrastructure (single-node and multi-GPU), with proper experiment tracking, reproducible configs, and the kind of logging that lets your future self understand what your past self did.
• Stay current with the literature: training methods, fine-tuning techniques, medical foundation models, multimodal architectures. Synthesize what's actually relevant for Atria rather than reporting every new arXiv preprint as a strategic threat.
• Collaborate with clinicians on what "good" looks like for each model, and with research partners on joint training and evaluation work.

Requirements

• A bias toward shipping models, not papers about models. You would rather have a working v1 in front of clinicians next month than a beautiful methodology that ships next year.
• A scrappy streak. You can pick up an unfamiliar fine-tuning technique, training framework, or clinical concept on a Wednesday and have a credible experiment running by Friday.
• A serious drive to keep getting better. You read other people's code, papers, training logs, and post-mortems. You treat being wrong as cheap information.
• Graduate degree (PhD preferred, Master's with strong research record) in computer science, machine learning, computational biology, biomedical informatics, or a closely related field — or a strong open-source track record in modern training and fine-tuning.
• Hands-on experience training and fine-tuning modern deep learning models, with a track record of shipped or published models you personally trained: 4+ years.
• Deep, current fluency with modern fine-tuning and post-training methods: SFT, PEFT (LoRA, QLoRA, adapters), preference tuning (DPO and successors), distillation, and continued pre-training.
• Strong working knowledge of the open-source model ecosystem: which models are state of the art, which are overrated, and what's worth fine-tuning for a given problem.
• Strong Python and PyTorch, with hands-on experience with the Hugging Face ecosystem (transformers, datasets, PEFT, TRL, accelerate) or equivalent training stacks.
• Practical experience with training infrastructure: distributed training, mixed precision, efficient data loading, experiment tracking (W&B;, MLflow, or similar).
• Discipline around evaluation and ablation: you treat benchmarking, calibration, and "what would this have looked like without that change?" as part of the modeling work.
• Genuine interest in healthcare and the responsibility that comes with building models that affect patient care.

Nice to have

• Experience building multimodal medical models, combining clinical text with imaging (radiology, pathology, ophthalmology), structured labs, or physiological signals.
• Familiarity with clinical/biomedical foundation models (e.g., Med-PaLM, MedGemma, BioMedLM, BiomedCLIP, RadFM) and the medical model literature.
• Peer-reviewed publications in ML, NLP, computer vision, or clinical informatics venues. • Experience with reinforcement learning, reasoning model training, or other frontier post-training techniques. • Experience with model interpretability and uncertainty quantification for clinical settings. • Prior collaboration with clinicians on models that shipped.

Salary range: $180,000 $260,000 base salary + performance-based bonus