Lock 2 · Design

Data Planning

Finding data you can actually use, handling it lawfully, and being honest about its gaps — the part of the blueprint most teams underestimate.

8 minute read · applies to Lock 2

A builder organising field-collected survey sheets beside a laptop with abstract chart shapes.

Inventory before you architect

List every piece of data your solution touches: training data, retrieval corpora, user inputs, generated outputs, logs. For each, answer four questions — where does it come from, do you have the right to use it, where does it live, and when is it deleted? This table is the backbone of both your data plan and your responsible-AI plan.

Where Indian builders actually find data

Government open data: data.gov.in, state portals, ministry dashboards
Sector systems: IMD and INCOIS (weather/ocean), Agmarknet (prices), eCourts (cases), UDISE+ (education), HMIS (health aggregates)
Research corpora: AI4Bharat (Indian languages), GBIF (biodiversity), open satellite imagery (Sentinel, Landsat)
Published documents: acts, rules, scheme guidelines — ideal retrieval corpora
Primary collection: your own pilot, with explicit consent — often your highest-quality option

Rights and consent are design constraints

Scraped personal data, leaked datasets and "found on a forum" corpora are disqualifying shortcuts — they fail the responsible-AI rubric and India’s DPDP Act framework. If your product collects personal data, consent must be specific, informed and revocable, and children’s data needs guardian consent. Design the minimum: every field you do not collect is a risk you do not carry.

Be honest about gaps — it scores better

Every Indian social-impact dataset has gaps: stale updates, English bias, urban skew, missing districts. Evaluators know this. A data plan that names its gaps and designs around them — local validation sets, uncertainty display, human fallback — scores higher than one that pretends the data is clean. "We tested on 200 locally collected samples because the public dataset skews North Indian crops" is a sentence that wins points.

Before you submit — the checklist

Data inventory: source, rights, storage, deletion for every item
At least one verified-accessible dataset per model feature
Consent flow designed for any personal data collected
Known gaps named, with mitigation
A local validation plan that does not depend on perfect public data

Related guides

Lock 2 · Design

Put this guide to work.

The rubric this guide addresses is waiting inside your team’s build journey.