Want to know if this job is worth applying to?
New York, New York, United States
Hybrid
We are building the first-ever bank for enterprise data. There are over 35 million businesses in the US alone, many of which are sitting on an asset they don’t know they have. At Replay, we help businesses monetize their historical operating data by licensing it to the frontier AI and data labs for model training. Replay sits at the intersection of advancing the frontier of AI, while helping companies generate millions in growth capital without having to give up equity, take on debt, or lay off employees.
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. The data does not arrive in one clean modality. It spans messages, documents, tables, files, images, metadata, and provider-specific structures, with important context distributed across all of them.
You will improve how well our system understands and protects that data. Your initial scope will be a prioritized subset of named-entity recognition, entity and identity resolution, structured extraction, classification, semantic review, or other model-backed parts of the de-identification pipeline. We do not expect one person to be an expert in every modality. The goal is measurable improvement in the areas you own: better precision, recall, F1, high-risk coverage, and preserved data utility across the failure modes that matter.
This is an applied, production-facing ML role. You will study errors, form hypotheses, build datasets and experiments, improve or replace models, and ship the result into a live pipeline. Evaluation, reproducibility, observability, and safe releases matter because they let us identify, ship, and verify meaningful model improvements in production.
Own and improve NER, entity resolution, structured or tabular detection, document understanding, semantic review, or related de-identification systems
Transform model failures and capability ceilings into a prioritized improvement roadmap
Design active-learning loops that combine model sweeps, LLM-assisted review, clustering, and uncertainty signals to identify the examples most worth hand-labeling
Build representative datasets and benchmarks, and use decision-relevant metrics to reveal strengths, weaknesses, uncertainty, and failure costs
Choose and combine deterministic rules, classical ML, fine-tuning, embeddings, multimodal models, and LLM-based approaches based on the problem and evidence
Design experiments, tune thresholds, analyze precision-recall and utility tradeoffs, and explain which changes are real, uncertain, or limited to particular conditions
Productionize improvements with reproducible artifacts, evaluation evidence, runtime instrumentation, and safe rollout
Optimize inference cost, latency, and throughput without hiding regressions in quality or high-risk recall
Build high-fidelity evaluation environments with seeded failure modes and programmatic verifiers that expose subtle regressions
Build reliable model- or agent-based harnesses with bounded behavior and explicit output verification when the problem calls for them
Partner with Applied Science on measurement and calibration, Data and Product Engineering on pipeline and review systems, and Security and Quality on acceptable risk
Use AI engineering tools deeply to accelerate research, implementation, error analysis, and evaluation while verifying their output
Model improvements generalize beyond the examples used to develop them and hold up in replay, shadow, and production evidence
Priority modalities and entity classes show credible improvements in precision, recall, F1, or other decision-relevant quality measures
High-risk misses decline without unacceptable over-redaction or loss of useful structure
New formats and modalities can be covered without relying on brittle one-off fixes
Improvements reduce meaningful delivery risk, review or rework burden, or loss of data utility rather than moving only an isolated benchmark
The team can explain why a model changed, where it improved or regressed across consequential failure modes and data segments, and whether the change should ship
The path from error discovery to a trustworthy production improvement becomes faster and more repeatable
Quality gains remain inside acceptable inference-cost, latency, and operational constraints
You have 3+ years of professional machine learning or software engineering experience, including improving models in production
You have startup experience, enjoy broad ownership, and thrive when requirements are evolving or incomplete
You use modern AI tools fluently and verify their output
You have personally moved model quality through error analysis, data work, experimentation, implementation, deployment, and iteration
You have a strong grasp of precision, recall, F1, calibration, thresholding, class imbalance, imperfect labels, distribution shift, and representative evaluation
You are an applied engineer first: a strong Python and software engineer who can work inside data pipelines and production systems, not only notebooks
You have a bias toward action while maintaining scientific and engineering rigor
You are curious and stay current with relevant state-of-the-art methods
You choose techniques based on the shape of the problem and can combine deterministic, statistical, neural, and LLM-based approaches
You communicate uncertainty and tradeoffs clearly to scientists, engineers, and people making delivery or risk decisions
You want to focus on research novelty without owning measurable production improvement
You prefer optimizing one aggregate benchmark without investigating consequential failure modes, data segments, and failure costs
You want data preparation, evaluation, deployment, and production diagnosis to belong entirely to other teams
You reach for a larger model before understanding the errors, constraints, and simpler alternatives
You do not want AI tools to be part of your daily engineering and research workflow
Experience with NER, entity resolution, information extraction, document understanding, multimodal systems, or privacy-preserving ML
Experience with hyperparameter tuning, data augmentation, model merging, ensembles, knowledge distillation, or multimodal model training
Experience fine-tuning or adapting transformer, GLiNER, embedding, vision-language, or small specialized models
Experience with active learning, uncertainty sampling, weak supervision, human-in-the-loop review, or LLM-assisted evaluation pipelines
Experience building goldens, adversarial corpora, replay systems, model bakeoffs, agentic harnesses, or programmatic evaluation environments
Experience with difficult ML or labeling problems
Experience with ONNX Runtime, TensorRT, model pruning, quantization, or other CPU/GPU inference optimization
Experience with sensitive enterprise data or other high-trust production systems
Experience with synthetic data generation and managing the synth-to-real gap
