Want to know if this job is worth applying to?
New York, New York, United States
Hybrid
At its core, Replay was founded to help founders. We started by supporting startups through shutting down, but we have since expanded into unlocking a new revenue stream for all types of businesses.
In 2025, we had a unique insight: the data every company generates each day through collaboration, communication, and building is some of the most valuable training data in the world. Public and synthetic data can only get frontier models so far, so the next generation of model progress depends on real, proprietary data grounded in how actual businesses operate. We are a primary source of it, partnering directly with the frontier AI labs building what comes next.
We have scaled from $0 to a multi-eight-figure run rate in a matter of months
We have raised from top-tier investors, including Floodgate, Afore, Ludlow, and Hustle Fund
We are small enough that you will carry outsized responsibility and grow as quickly as the company does
You will partner with and build for some of the fastest and most important companies in the world
You will help build a massive, category-defining business from the ground floor
Replay turns sensitive internal enterprise data into de-identified datasets without destroying the structure and meaning that make the data valuable. That creates a difficult measurement problem. A system can improve aggregate F1 while missing a high-risk slice, remove more sensitive information while also destroying useful context, or pass one stage while defects escape somewhere else in the pipeline.
As Replay's first Data Scientist focused on evaluation, you will establish how we know whether that data is actually getting better. You will build the datasets, experiments, quality measures, and feedback loops that expose hidden failures, accelerate model and pipeline improvement, and give the team confidence in what it delivers.
This is a hands-on, zero-to-one role at the intersection of data science, AI, and a real production system. You will write Python and SQL, construct evaluation corpora, study failure patterns, design comparisons, calibrate human and model-based judgments, and turn the result into a clear decision. The questions are scientifically difficult, but the output must be practical enough to change what the team builds and ships.
You will work closely with Machine Learning, Product Engineering, Data Engineering, Security, Quality, domain experts, and the team making delivery decisions. Machine Learning Engineers own changing model behavior. You own the credibility of the evidence used to decide whether a model, pipeline, or delivery change actually made the data safer or more useful.
Did a higher NER or entity-resolution score actually reduce sensitive misses across the messages, documents, tables, and providers that matter?
Is a new model finding more sensitive information, or simply removing more of the useful structure our customers need?
Can we trust a golden dataset, a human review process, or an LLM judge enough to use it for a release decision?
Which customer, modality, entity, language, or format slices are hidden by a strong aggregate result?
Where did a quality loss enter between source data, processing, de-identification, review, and delivery?
What is the smallest credible experiment that would tell us whether to ship, revise, or stop a change?
Define what high-quality and safe-to-deliver data mean across de-identification, structure preservation, semantic coherence, and customer utility
Design representative samples and build golden, adversarial, replay, and production-like corpora with explicit provenance, labeling policy, agreement, adjudication, and versioning
Turn ambiguous concepts such as “useful,” “clean,” or “safe” into measurable claims with known uncertainty and clear decision consequences
Evaluate detectors, models, prompts, judges, thresholds, review workflows, and pipeline changes using comparisons that can support a real decision
Break aggregate results into the modalities, providers, entity classes, customer contexts, languages, formats, and risk tiers that reveal consequential failures
Connect local measures to escaped sensitive information, avoidable over-redaction, preserved data utility, review burden, rework, and delivery acceptance
Build reproducible analysis, evaluation pipelines, and high-fidelity environments using Python, SQL, synthetic data, historical replay, seeded failures, and programmatic verifiers
Establish holdout and evaluation practices that keep the evidence trustworthy while model and product teams iterate quickly
Use modern AI tools deeply for analysis, corpus development, coding, review, and hypothesis generation while independently verifying their output
The team has a decision-grade baseline for a priority Clean Data quality claim and trusts it enough to use in model, pipeline, release, and delivery decisions
Improvements are judged by the slices and failure costs that matter, not only by an aggregate benchmark
The company can distinguish a true gain from label noise, sample bias, leakage, evaluator error, or a shifted workload
Changes that improve one stage cannot hide escaped defects, over-redaction, utility loss, or review burden somewhere else
At least one consequential decision changes because the evidence reveals a risk, tradeoff, or opportunity that was previously unclear
Evaluation becomes faster and more repeatable without sacrificing independence or rigor
Quality claims communicate uncertainty honestly and remain understandable to engineers, customers, and risk owners
You have at least three years of professional experience in applied science, data science, machine learning, quantitative research, or a closely related role
You have designed evaluations or experiments that changed a product, model, release, or operational decision
You understand sampling, uncertainty, precision, recall, F1, calibration, agreement, class imbalance, distribution shift, and imperfect labels
You can investigate messy, multi-stage data systems and determine where an apparent gain or loss actually came from
You are comfortable writing Python and SQL and building reproducible technical artifacts rather than handing requirements to an engineering team
You can protect the independence of an evaluation while collaborating closely with the people whose work it evaluates
You have startup experience and enjoy broad ownership, changing context, and building the measurement foundation while decisions are already moving quickly
You use AI tools fluently but do not confuse an articulate model output with valid evidence
You communicate uncertainty and difficult findings directly, without hiding behind false precision
You want to optimize models as your primary job rather than determine whether changes actually improve delivered data
You prefer descriptive dashboards that stop short of changing a decision
You treat labels, benchmarks, or model-based judges as ground truth without investigating how they fail
You need a perfectly defined dataset and research plan before you can make progress
You are uncomfortable disagreeing with a technically strong team when the evidence does not support its conclusion
You do not want AI tools to be part of your daily scientific and technical workflow
Experience evaluating NER, entity resolution, information extraction, document understanding, multimodal, retrieval, or LLM systems
Experience with privacy, de-identification, data quality, model risk, safety, or other high-trust decision systems
Experience designing human-review, adjudication, weak-supervision, or active-learning systems
Experience building adversarial corpora, replay systems, simulation environments, programmatic verifiers, or model-judge evaluations
Experience connecting offline measures to escaped defects, customer outcomes, review effort, or preserved data utility
Experience measuring quality across multi-stage batch or data pipelines
