Applied ML Scientist

ABOUT SAIVA AI

SAIVA AI applies machine learning to make optimal use of electronic health data for the most vulnerable healthcare population. Our mission is to improve patient outcomes by augmenting clinical decision-making with the power of AI. 

Based in Silicon Valley, our team is a group of passionate healthcare technology veterans, engineers, and data scientists, including from Stanford University, leveraging cutting edge technology to predict patient risk and provide tools that drive timely intervention. 

POSITION DESCRIPTION

As an Applied ML Scientist, you will use your extensive experience to build and improve models across structured + unstructured clinical data, pairing boosting models (like xgboost) with language models (LLM APIs, smaller Transformers you will fine tune or train) to lift real-world impact. You’ll prototype fast, run rigorous evaluations, and deliver sound research code that our MLEs will harden & optimize and our MLOps will ship & monitor.

RESPONSIBILITIES

In this position you will:

• For a given task, ship a prompted LLM baseline → fine-tuned small LM or BERT-class model, comparing precision@recall, cost, and latency.
• Build light RAG/explainability: return the strongest supporting sentence/span from notes that justifies the classification.
• Leverage domain NER (don’t rebuild it) to target specific acute phenotypes and improve downstream classifiers.
• Distill/quantize to hit latency/cost SLOs while preserving recall; hand off clean checkpoints/configs/eval harnesses.
• Work hands-on in Python, Jupyter, Git, SQL, MLflow; AWS (EC2/ECS/S3) or equivalent cloud.

Requirements

• 2+ years in industry doing hands-on NLP (IC), across prompting, fine-tuning, and training BERT-class models.
• Evidence of high-recall use cases with calibration and defensible thresholds; comfort with noisy note text.
• Undergrad degree required in Computer Science or Computational Linguistics (strong quantitative background).
• Nice to have: visible healthcare NLP work (papers, preprints, GitHub).