Synthetic dataset generation for AI projects

Solveion is an independent AI consultancy. We build custom AI systems for clients: assistants, automations, and retrieval systems. Those systems need training and test data before real client data exists, or where real data cannot leave the client. I need someone to generate that data. This is a developer role, not a labeling role. The work varies by project, but it looks like this: - Generate synthetic records to a schema I give you: support tickets, invoices, contracts, intake forms, chat transcripts - Build the generation pipeline in Python, with a seed and a config so the same run reproduces - Write matching evaluation sets: the input, the correct answer, and the reason it is correct - Build in the messiness on purpose. Typos, missing fields, contradictions, and edge cases, because clean data teaches a system nothing about production. - Document what you generated, how, and what the distribution looks like Volume and cadence change project to project. Right now I expect around 40 hours a month across two or three projects. What I need from you: - You have generated synthetic data before, whether with LLMs, Faker, SDV, or your own scripts, and you can explain the trade-offs between those - You write Python I can read and hand to someone else - You understand why a synthetic set that is too clean or too uniform is worse than no set at all - You will sign an NDA. Schemas and business rules come from client work. Useful but not required: experience with retrieval and RAG evaluation, differential privacy, or regulated data in finance, legal, or healthcare. Everyone starts with one paid pilot: generate 200 to 500 records against a schema I provide, plus a short note on your approach. Answer these in your proposal: 1. Describe one synthetic dataset you generated. What was the schema, how did you generate it, and how did you check it was any good? 2. I need 5,000 synthetic support tickets that will train a classifier. How do you stop the model from learning your generator's habits instead of the real problem? 3. Your hourly rate and monthly availability.