Applied Data Scientist, Health AI Evaluation & Datasets

Innodata (Nasdaq: INOD) is a global data engineering company. We believe that data and Artificial Intelligence (AI) are inextricably linked. Our mission is to enable the responsible advancement of artificial intelligence by providing the data, evaluation frameworks, and human expertise required to build AI systems that can be trusted at scale. We provide a range of transferable solutions, platforms, and services for Generative AI / AI builders and adopters. In every relationship, we honor our 36+ year legacy delivering the highest quality data and outstanding outcomes for our customers.

Scope of the Role:

Healthcare is one of the highest-stakes domains for generative AI. Clinical accuracy, patient safety, regulatory compliance, health equity, auditability, and workflow fit are the bar for shipping anything real. Innodata partners with foundation model labs, medical AI startups, payers, providers, pharma, and digital health companies building LLMs, multimodal systems, and AI agents for healthcare and life sciences.

As an Applied Data Scientist, Health AI Evaluation & Datasets, you own the design, measurement quality, and clinical validity of datasets used to train, fine-tune, and evaluate health-domain models. You bring clinical or biomedical fluency and data science rigor: you can read a clinical guideline, payer policy, medical literature artifact, or patient communication workflow; translate it into a measurable dataset and evaluation plan; and defend the methodology to sophisticated clinical, data science, and ML stakeholders.

You will work in a tight pod with a Technical Solutions Architect, Applied Research Scientist, AI/ML Research Engineer, and Language Data Scientists. Your role is to make sure the data, rubrics, review workflows, and measurement evidence are clinically realistic, statistically defensible, compliant, and useful for evaluation and post-training.

What You’ll Own:

• Translate customer goals — such as improving differential diagnosis, evaluating a clinical note summarizer, testing a RAG-based medical literature assistant, or creating preference data for patient-facing chatbots — into dataset specifications, taxonomies, rubrics, sampling plans, and acceptance criteria.
• Make multimodal health AI a core focus: design training and evaluation datasets across clinical text, medical images, waveforms, structured EHR data, claims, trial data, medical literature, patient communications, payer policies, drug information, and other clinical artifacts, as well as use cases such as clinical reasoning, medical QA, note summarization, medical coding, patient communication, utilization management, and literature synthesis.
• Design evaluations for retrieval-augmented and source-grounded health AI systems, including evidence citation, faithfulness, contraindication handling, guideline adherence, source freshness, and failure modes caused by incomplete, conflicting, or stale context.
• Define sampling strategies, label schemas, inter-annotator agreement targets, adjudication workflows, SME review patterns, and quality thresholds in partnership with Language Data Scientists, clinicians, biomedical experts, and quality teams.
• Build statistical and ML checks that make healthcare datasets trustworthy: stratified sampling across specialties and patient subgroups, bias and representation analysis, leakage detection, distribution shift checks, uncertainty estimates, reliability metrics, and subgroup performance analysis.
• Partner with Applied Research Scientists and AI/ML Research Engineers to instrument datasets into evaluation and post-training pipelines, including rubric-grounded LLM-as-judge prompts, regression suites, model comparison workflows, experiment tracking, and model-improvement feedback loops.
• Evaluate health AI behavior beyond surface accuracy: calibration, hallucination on safety-critical content, refusal appropriateness, robustness under ambiguity, equity across patient subgroups, and safe handoff in agentic or workflow-integrated systems. Reason concretely about clinical workflow fit: where outputs enter care delivery, what evidence a clinician or reviewer would need to trust them, when uncertainty must be surfaced, and how patient-facing, clinician-facing, payer, pharma, and operational use cases differ in risk.
• Own data quality from source intake through delivery, including de-identified clinical text, medical literature, synthetic cases, structured records, client policies,…