By Arun Balaji ·

The question

Mula AI is designed to understand local places through spatial perception. Before it can perceive a region, it has to know which facts about that region it can trust, how precise they are, and how long they stay true.

In July 2026 we built a research bundle for a regional agro observatory covering South India to answer a practical question: which official agricultural data can a small, local AI actually rely on?

Place and method

  • Region: South India (Kerala, Tamil Nadu, Karnataka, Andhra Pradesh and Telangana), with Kerala as the pilot and Idukki district as the first focus.
  • Sources: 14 official sources, covering state and national agricultural statistics, commodity boards, district weather advisories, remote sensing, soil, water and administrative boundaries.
  • Checks: each source was assessed for official status, units, reference period, estimate status, licence, and whether it could be retrieved in a stable, dated form.
  • Verification: every number we kept was checked against its official publication. An earlier automated research report was set aside because it misread the brief; none of its agricultural claims were used.
  • Reproducibility: the bundle includes schemas, valid and deliberately invalid test cases, and validators, and it passes its own complete validation.

What we found

  • Only 2 of the 14 sources survived as facts we could keep: Kerala's Agricultural Statistics 2022-23 and Tea Board India's 2024 tea production figures, which are provisional. Eight became references, one became a future live feed, and three were deferred.
  • No official crop boundary and no current field signal survived. The district weather advisory had no fixed, dated bulletin we could keep, and the remote-sensing, soil and water layers lacked a stable source, licence or precision we could verify.
  • Government data is not automatically open data. Licences had to be checked source by source. Where terms were unclear, we kept only the facts, with attribution, and never the source files.
  • Official sources disagree in ways that matter. Kerala's statistics report 65,980 tonnes of tea for 2022-23 (final). The Tea Board reports 58.38 million kg for 2024 (provisional). Different years, units and status, so they are shown side by side and never blended into a trend.
  • Regional facts describe context, not conditions. Idukki had 25,315.33 hectares of tea and 364.19 hectares of paddy in 2022-23. That is a useful picture of the district, but it says nothing about any particular estate today.

What this means for Mula AI

The research produced a simple grammar for how Mula AI should describe what it knows.

  • Truth: whether something was observed, reported, derived, inferred, hypothesised, unknown or on hold.
  • Precision: national, state, district, derived region, estate context or field evidence. A fact never becomes more precise than its source.
  • Freshness: static, annual, seasonal, weekly, daily or live. Advisories expire, and expired data is history, never current.

Regional context never upgrades field truth. For a planned paddy survey in Tamil Nadu, only the drone survey itself, with its location, accuracy and ownership records, can describe the fields.

What remains open

  • Admitting a dated, verifiable district weather advisory.
  • Licensed district boundaries from the Survey of India, in place of approximate markers.
  • Stable remote-sensing, soil and water layers.
  • The Tamil Nadu paddy drone survey, which is planned but not yet carried out.

Sources & references

All research & insights