Technical Program Manager - Frontier AI Data Labeling

About the Role

Frontier models are only as good as the data they learn from. At Hippocratic AI, that data decides whether an AI agent safely guides a patient through a medication change at 2 a.m.

Our data powers Polaris, a constellation of models with a 700B-parameter core and specialized safety supervisors. In our evaluations, Polaris reaches 99.95% accuracy on drug-safety benchmarks versus 87.9% for the closest voice-deployable frontier model. The labeling, evals, and rater calibration behind it are how we earn and defend numbers like that.

We’re hiring a Technical Program Manager to run the human-in-the-loop data labeling engine behind our frontier healthcare AI. That spans audio and conversation annotation, clinician-labeled training data, preference and RLHF data, and the evaluations that decide what ships. You’ll work daily with ML researchers, applied scientists, and licensed clinicians. The data you produce feeds directly into our model training and evaluation loops.

What You’ll Do

• Run the data pipelines behind our models. Own end-to-end delivery of labeling and annotation programs, including audio and voice-conversation labeling, supervised labeling, and preference and RLHF data. You’ll manage scoping, task design, execution, QA, and delivery against live model-development cycles.

• Turn evals into products. Work with ML and Applied Science to turn evaluation and training-signal needs into concrete labeling tasks, rubrics, and success criteria. Your evals become the benchmarks that gate releases.

• Build and scale an expert rater workforce. Source, screen, and onboard clinicians and specialized raters through marketplaces, vendors, and in-house teams. This includes designing qualification tasks, setting rates, and scaling pools up and down as programs change.

• Write the guidelines that shape model behavior. Author and iterate labeling guidelines and annotation rubrics, especially for subjective, high-stakes clinical judgment.

• Engineer data quality at scale. Design QC systems using gold sets, calibration sessions, spot checks, and inter-rater agreement tracking. Catch rater drift and annotation bottlenecks before they reach training.

• Drive velocity and accountability. Track throughput, cost, and quality. Manage vendor relationships (SOWs, performance reviews), choose labeling and QA tooling, and keep Engineering, Clinical, and Infrastructure aligned on dependencies and timelines.

What Success Looks Like

• By Day 90: You own a live data labeling or eval program through a major delivery milestone. You’ve cleared a critical bottleneck in the pipeline and shipped one measurable quality or throughput improvement.

• By 12 months: You run multiple concurrent labeling and eval programs. Your calibrated rater bench is trusted by ML teams, and you can trace model improvements back to the data you delivered.

What You Bring

• 5+ years in technical program management or data operations in an AI/ML or data-intensive environment (3+ years considered for exceptional candidates)

• 2+ years hands-on with data labeling, annotation, or model evaluation, including writing guidelines and managing quality at scale

• Experience sourcing and managing contractor or expert annotator workforces through marketplaces (e.g., Mercor, Upwork, Toloka) or labeling vendors

• Working knowledge of how annotation quality affects model training, including preference data, pairwise comparisons, and rater calibration (RLHF)

• A strong grasp of quality metrics (IAA, precision/recall on gold sets, throughput vs. accuracy tradeoffs)

• A track record of owning critical data or eval projects under time pressure, working with ML engineers and applied scientists, and driving results without direct authority

• Comfort with ambiguity in a fast-moving, healthcare-adjacent AI environment

Nice to Have

• Experience running or supporting RLHF or other training-data pipelines at an AI lab or applied ML team

• Clinical, healthcare, or other regulated, high-stakes data experience

• Familiarity with Label Studio, Scale AI, Surge AI, or SageMaker Ground Truth

• Experience qualifying specialized raters such as clinicians or bilingual SMEs

• Exposure to reward hacking and specification gaming, and how annotation design can mitigate them