Synthetic data generation for Indian companies

By Vishal · published 25 October 2026

When synthetic data is the right tool

Four situations come up again and again in Indian companies, and synthetic data solves each of them differently.

It is the wrong tool when the real data itself is wrong. A synthetic copy of a biased dataset is a bigger biased dataset. The feasibility note says this plainly if we find it.

What we generate

• Tabular data — customer, transaction, claims, inventory and sensor tables with the column distributions, correlations and business rules of the original (a customer's city matches their PIN code; a refund never exceeds the order). Generated with conditional tabular models or, for small tables, calibrated statistical models that are easier to audit.
• Text — support tickets, product reviews, clinical or legal notes, chat transcripts in English and Indian languages, generated from your taxonomy and tone with a language model and filtered against your real corpus for leakage.
• Images — product photos, document scans (invoices, KYC forms, prescriptions) and defect images for vision models, with annotations produced alongside so labelling cost drops to review time.
• Time series — demand, footfall, energy and sensor sequences that keep seasonality, trends and festival spikes.

How we measure fidelity and privacy

Every delivery ships with a one-page scorecard. Without it you are trusting a feeling.

Fidelity: per-column distribution distance against the real data; pairwise correlation difference; and the test that matters, train-synthetic-test-real — a model trained only on synthetic data, scored on held-out real data, against the same model trained on real data. If the gap is more than a few points, the synthetic set is not good enough yet.

Privacy: nearest-neighbour distance between synthetic and real records (no synthetic row should be a near copy of a real one), membership-inference tests, and a check that rare combinations which identify a person do not reappear. Where the risk warrants it we generate with differential privacy and state the budget.

Utility for your task: the only score that finally counts. We agree the downstream metric in week one and report it at the end.

How a project runs

1. Feasibility note (free) — a call, a look at the schema and a sample (or its profile if you cannot share rows), and a one-page note: what to generate, how to measure it, pilot price.
2. Two-week pilot (fixed price) — one table or one text type, the scorecard, and a go/no-go.
3. Full generation and handover — all agreed data types, the generator packaged so your team can regenerate on a schedule, documentation, and a review of how the data will be used.

Work happens inside your cloud account or on an isolated machine in our Bengaluru lab; we do not keep copies of your real data after handover.

Synthetic data and the DPDP Act

India's Digital Personal Data Protection Act, 2023 governs personal data. Synthetic data that cannot be linked to a real person is a strong way to reduce what you hold and share, but it is not a magic word: a poorly generated set can leak, and the privacy tests above exist for that reason. We document the generation method and the privacy results so your data-protection officer or lawyer can assess it. This is engineering support, not legal advice.

What you receive

• The synthetic datasets, in your formats, with a data dictionary.
• The fidelity and privacy scorecard and the notebooks that produced it.
• The generator as a repository you own, with instructions to run it again.
• A short written recommendation on where synthetic data should and should not be used in your company.