Synthetic Data Engineer Job Description

A Synthetic Data Engineer builds the systems that manufacture artificial records, images, text and sensor streams for training and testing machine learning models. The output has to behave like real data wherever the model is sensitive to it, and it has to stop short of reproducing any real person's record. Holding both requirements at once is the whole job.

Work starts with a request from a model team. A fraud group wants transaction histories with extra examples of rare attack patterns. A perception team wants camera frames of a warehouse aisle under lighting it never recorded. A health analytics group wants patient-level tables it can hand to an outside vendor without exposing anyone. The engineer turns each request into a specification: which fields, which relationships between fields must survive, how many rows or frames, which edge cases need extra coverage, and what privacy guarantee the release must carry.

Next comes the choice of generator. For tabular data, open-source libraries such as the Synthetic Data Vault (SDV) learn statistical patterns from a training table and then sample new rows that imitate them. For text and conversational data, LLM-driven tools such as NVIDIA's NeMo Data Designer take user-supplied model endpoints and handle batching, parallelism and validation of the generated records. For vision and robotics, simulation frameworks such as NVIDIA Omniverse Replicator render physically accurate 3D scenes; NVIDIA's documentation pitches the tool for long-tail anomalies and edge cases that are impossible to collect in the real world. Many pipelines mix methods, starting from simulated ground truth and then using a generative model to vary textures, weather and backgrounds.

Evaluation takes as much care as generation. Engineers measure fidelity (do the marginal distributions and correlations match the source?), utility (does a model trained on the synthetic set hold up on a real holdout?) and privacy (could an attacker link a synthetic row back to a real one?). A dataset can pass a visual sniff test and still drop the rare cases a model needs, so these checks run on every build rather than once at launch.

Privacy engineering sits under all of it. When a release claims differential privacy, the engineer sets and tracks the privacy parameter, epsilon, and documents what the chosen setting costs in accuracy. For regulated records, the engineer works with legal and compliance staff on whether a synthetic release counts as de-identified under the rules that apply to that data.

Day to day, the role looks like data engineering with a statistician's habits: orchestration jobs in Airflow or Dagster, GPU batch runs, dataset versioning, and a reviewer's eye for generated samples that look plausible but quietly encode a label leak. The engineer usually sits in an ML platform, data or research group and may be the person who signs off that a synthetic dataset is fit to train on.

Education. The usual starting point is a bachelor's degree in computer science, statistics, mathematics or a related quantitative field. BLS describes the same entry path for data scientists, the occupation that covers statistical modeling work, and notes that some employers require or prefer a master's or doctoral degree. For synthetic data roles, a graduate degree matters most where the job involves designing new generative models or formal privacy mechanisms rather than running established ones.

Experience. As a rough guide, mid-level hires bring 3 to 5 years in data engineering, ML engineering or applied statistics, including at least one project that generated and evaluated a synthetic dataset end to end. As a rough guide, senior engineers bring 6 or more years and own the evaluation framework and the privacy sign-off.

Technical skills.

• Python data work: pandas, NumPy and PyTorch, plus solid SQL against warehouse tables.
• Generative methods: copula and Bayesian network synthesizers, CTGAN and TVAE for tables, diffusion models for images, and prompt-and-validate LLM pipelines for text.
• Evaluation: distribution distance metrics, train-on-synthetic, test-on-real benchmarks, and membership-inference and linkage attacks run against your own output.
• Privacy: differential privacy accounting with libraries such as OpenDP or Opacus, plus classic de-identification methods (suppression, generalization, k-anonymity).
• Pipelines: orchestration, containers, GPU scheduling and dataset versioning, so every model run can name the exact dataset build it trained on.
• Simulation, for vision and robotics roles: 3D tools such as Isaac Sim, Unreal Engine or Blender, domain randomization and sensor noise models.

Reading list. Two federal documents are worth reading closely. NIST SP 800-188, De-Identifying Government Datasets, sets out the choice among publishing de-identified data, publishing synthetic data based on identified data, offering a query interface, or sharing inside a protected enclave. NIST SP 800-226 explains how to evaluate a differential privacy claim and catalogs the privacy hazards, common pitfalls that appear when the math is turned into software.

Credentials. There is no certification built for synthetic data engineering. Some engineers add the Certified Information Privacy Technologist (CIPT) from the IAPP to show privacy engineering grounding, or a cloud ML credential such as the Professional Machine Learning Engineer certification from Google Cloud. A portfolio that includes an evaluation report, and not just a generator, shows the core skill directly.

Working style. The job rewards people who can explain to a product manager or privacy officer why a tighter privacy setting produces a less accurate model, and who write documentation an auditor can follow months later. Comfort reading research papers helps, since generator and evaluation methods come out of conference work.

BLS does not track synthetic data engineers as a separate occupation, so the clearest federal signal comes from data scientists, the occupation that covers the statistical modeling at the center of this work. BLS projects employment of data scientists to grow 35 percent from 2025 to 2035, much faster than the average for all occupations, with about 24,800 openings a year on average. BLS ties part of that demand to firms continuing to integrate artificial intelligence systems into their workflows. Those figures describe the whole occupation; BLS does not report synthetic data specialists separately.

Pay follows the data scientist band. In May 2025, BLS reported a national median of $120,230 for data scientists, with the 10th percentile at $67,240 and the 90th percentile at $199,130. In software publishers, one of the industries BLS lists for the occupation, the median was $156,220.

Platforms and vendors. In March 2025 Nvidia reportedly acquired Gretel, a San Diego synthetic data startup; TechCrunch reported the price was said to be nine figures and that Gretel's team of roughly 80 employees would be folded into Nvidia. NVIDIA now documents NeMo Data Designer as an orchestration framework for generating high-quality synthetic data. When generation comes from a vendor service, the engineer's own contribution is the specification, the evaluation and the governance wrapped around it.

Model collapse research. A paper published in Nature on July 24, 2024 found that indiscriminate use of model-generated content in training causes irreversible defects, in which the tails of the original content distribution disappear; the authors called the effect model collapse. A 2025 paper in the ICML proceedings narrowed the picture: replacing all real data with successive generations of synthetic data collapsed in every setting the authors studied, while accumulating synthetic data alongside real data kept models stable. The practical lesson for this job is about provenance and mixing: knowing which training rows are synthetic, which generator made them, and which build they came from.

Federal privacy guidance. NIST published SP 800-226, Guidelines for Evaluating Differential Privacy Guarantees, in March 2025. It lists synthetic data generation among the tasks differentially private algorithms exist for, and it says differential privacy is still making the transition from theory to practice. NIST's Privacy Engineering Program also runs the Collaborative Research Cycle, a benchmarking exercise with over 500 de-identified excerpts, including differentially private synthetic data, each scored with the SDNist library for fidelity, utility and privacy. Its larger benchmark set has 130 features and 161K records. The Census Bureau has public experience as well: its SIPP Synthetic Beta made linked survey and administrative data available as synthetic microdata, and a 2024 Census working paper on that product discussed a validation system in which the data holder runs an outside researcher's code on the internal data.

Where the career goes. From this role, paths lead toward ML platform leadership, privacy engineering, or applied research on generative models and evaluation methods. The documentation and certification side of the work also maps onto AI governance and assurance jobs, where someone has to show what a model was trained on.

Dear Hiring Manager,

I am applying for the Synthetic Data Engineer position on your ML platform team. For the past four years I have worked as a data engineer at a regional health insurer, and for the last two I have owned the pipeline that produces synthetic claims tables for our fraud and care-management model teams.

When I took the project over, the synthetic tables were generated by a single script with no evaluation step. Model teams trusted the output because it looked right in a spreadsheet. I rebuilt the pipeline as an orchestrated job with three gates: a fidelity report comparing distributions and pairwise correlations against the source, a train-on-synthetic, test-on-real check run by the model owners, and a linkage attack that tries to match synthetic members back to real ones. A build that fails any gate stops and pages me instead of publishing.

The privacy gate changed how the business uses the data. Our compliance office used to review each release by hand. Now every release ships with a short memo that states the generator, the privacy setting, the attack results and the known gaps, and compliance reviews the memo instead of the rows. That memo format is the piece of work I am proudest of, because people outside engineering can read it.

I am drawn to your team because you generate text and tabular data for outside customers, which raises the bar on documentation and on catching quiet failures like label leakage. I would bring a habit of treating evaluation as part of the product, plus a working knowledge of differential privacy accounting that I have applied in production.

I would welcome the chance to walk you through one of my release memos.

Sincerely, Jordan Ellis