Granted patent · Future AGI · Issued 21 Apr 2026 · filed 27 May 2025 · USPTO · US 12,608,610 B1
Synthetic data generation
How do you produce realistic data for training and testing AI systems when the real data is scarce, or off-limits?
What I did
Co-inventor, with Nikhil Pareek, Rishav Hada and Srikanth Malyala. Assigned to Future AGI Inc.
Inventors
Nikhil Pareek, Rishav Hada, Srikanth Malyala, N.V.J.K Kartik
Full title
Synthetic data generation system and method
In plain words
You describe the data you need. The system turns that into a schema and generates samples that follow it: the fields, how they relate, and how values should be distributed.
It supports three starting points: nothing at all (seedless), a few real examples (seeded), or examples plus a knowledge base.
How it works
- Claim 1: receive domain requirements and a scenario type (Seedless, Seeded, or Seeded + Knowledge Base); define a structured schema with data fields, relationships and distributional targets; generate an initial sample set with a neural template-driven model trained on domain-specific data; and continue from there.
What it established
- A granted US patent, application 19/219,261.
Where it stops
- A patent is a legal record of an invention. It is not evidence that the method beats alternatives. There are no benchmark claims here.
Receipts
- USPTO Official Gazette — US 12,608,610 B1Title, inventors (Kartik fourth of four), assignee Future AGI Inc., filed 27 May 2025, issued 21 Apr 2026, claim 1.
Next, a little further back in timeAI Evaluation 1.0 →