Research Engineer - Data Quality & Evals

Epsilon Health; San Francisco (CA); On-site; USD 120,000 - 180,000; Full time

Job summary

Epsilon Health in San Francisco seeks a Research Engineer to own data quality and evaluation across ML efforts for radiology report generation. You will build filtering and curation pipelines, create quality signals, and collaborate with radiologists to define criteria.

You will design evaluation methods beyond surface text, measure clinical accuracy, and ship scalable tooling for data, training, and evaluation.

Qualifications

• 2+ years of ML, data engineering, or related experience.
• Strong Python and software fundamentals; able to build tooling and pipelines from scratch.
• Willing to grow in data quality and evaluation areas; comfortable in a fast-moving research setting.

Responsibilities

• Build data filtering and curation pipelines to keep training and evaluation sets clean.
• Develop model-based data quality signals and automated flagging.
• Collaborate with radiologists and annotators to define quality criteria and edge cases.
• Design evaluation methodology for radiology report generation, measuring clinical accuracy and style.
• Create and maintain clinical benchmark sets and regression tests.
• Build evaluators (LLM-as-judge, rubric grading) and track offline vs production outcomes.
• Ship tooling across data, training, and evaluation to speed research.

Skills

Python

Data pipelines

ML engineering

Agency/initiative

Tools

Spark

Airflow

Databricks

Job description

About Us

We're tackling one of healthcare's most critical challenges in medical imaging and diagnostics. Our company operates at the intersection of cutting-edge AI and clinical practice, building technology that directly impacts patient outcomes. We've assembled one of the industry's most comprehensive and diverse medical imaging datasets and have a proven product-market fit with a substantial customer pipeline already in place.

Role Overview

We're seeking a Research Engineer to join our ML Research team, owning data quality and evaluation across our modeling efforts. On the data side, you'll build the filtering and curation systems that keep our VLM and classifier training sets clean - catching label noise, misaligned image-report pairs, duplicates, and quality issues at scale. On the evaluation side, you'll design how we measure radiology report generation, incorporating clinical accuracy, completeness, hallucination, and reporting style at once. The signals you build feed directly into how our foundation-model and post-training teams train and improve their models. This is a broad, dynamic role for someone with the agency to identify where the research team is bottlenecked and address it directly.

Key Responsibilities

• Build data filtering and curation pipelines that keep VLM and classifier training sets clean, detecting label noise, misaligned image-report pairs, duplicates, corrupted studies, and low-quality samples at scale.

• Develop model-based data quality signals (alignment scoring, automated flagging, active-learning loops) to surface the ambiguous or high-value cases worth human review.

• Partner with radiologists and annotators to define quality criteria, adjudicate edge cases, and turn clinical judgment into reusable, scalable filters.

• Design evaluation methodology for report generation that goes beyond surface-level text overlap, measuring clinical accuracy through entity and relation extraction, hallucination and omission rates, and adherence to reporting style.

• Build and maintain clinical benchmark sets, stratified by modality, pathology, and difficulty, with rigorous attention to train/eval contamination.

• Develop and validate model-based evaluators (LLM-as-judge, rubric grading) against radiologist judgment, and track how offline eval correlates with production and clinical outcomes.

• Build continuous evaluation and regression testing so the team can measure every model change quickly and trust the result.

• Work across the research stack (data, training, and evaluation) finding bottlenecks and shipping the tooling that lets research scientists move faster.

Qualifications

• 2+ years of industry or research experience in ML, data engineering, or a related area

• Strong Python and solid software engineering fundamentals; comfortable building tooling and data pipelines from scratch

• Strength in one or both of our core areas, with the willingness to grow into the other:

  • Data quality: dataset curation, filtering, deduplication, label-noise detection, or data-centric ML

  • Evaluation: designing metrics or eval harnesses for generative models, LLM-as-judge, or NLG / factuality evaluation

• Demonstrated agency, i.e. a habit of identifying important problems and driving them to a result without waiting to be told

• Comfort working in an ambiguous, fast-moving research environment and collaborating closely with research scientists

Preferred Qualifications

• Experience with medical imaging or clinical data (DICOM, radiology reports, clinical NLP)

• Familiarity with vision-language models or multimodal training

• Experience building human-in-the-loop annotation or review workflows, and reasoning about inter-annotator agreement

• Experience with clinical accuracy metrics for report generation (e.g., entity / relation extraction, RadGraph-style scoring)

• Experience with data pipeline and experiment tooling (Spark, Airflow, Databricks, or similar)

• Publications or open-source contributions in data-centric ML, evaluation, or medical AI