A Pakistani smallholder has no credit file. Most scoring models solve that by guessing. The Agri Credit Score doesn't — it scores what satellites and registries can verify on day one, leaves the rest at zero weight until it's genuinely observed, and tells the lender which is which.
We publish the approach openly. The scoreboard is free; the score is the product.
Ask most agri-scoring tools what a score is based on, and a large share of the weight rests on things that don't exist for a first-time borrower — transaction history, repayment record, platform behaviour. For a farmer who has never borrowed formally, those fields are empty, so the model quietly imputes them as "average". A bank's model-risk team spots that immediately, and rightly discounts the whole score.
Our earlier internal version had exactly this flaw: 42% of its weight sat on behaviour we cannot observe before our marketplace is live. ATI v1 inverts the architecture rather than papering over it.
Sentinel satellite archives are retroactive. The moment we capture a plot boundary, we can compute roughly three years of NDVI and radar history for that land — immediately, without waiting.
Gates are pass/fail, not points. Fail one and the file is referred for human review — it does not receive a weaker score. This mirrors EU checks-by-monitoring and Brazil's CPR central-registry anti-double-pledge check.
The day-one score — ATI-0 — is roughly 45% satellite evidence and 55% verified registry and declared data. Every input is obtainable for a Pakistani farmer this quarter: NADRA, land records, a boundary walk, and the Sentinel archive. Nothing in it is aspirational.
Behavioural signals — marketplace conduct and repayment history — sit at zero weight. Not estimated. Not imputed. Zero, until they are actually observed.
Every score carries a maturity tier, so a lender always knows which regime produced it. As real observations accumulate, dormant signals absorb weight from the proxies. At origination this follows the MYbank pattern (satellite and registry can underwrite a farmer with no financial history — 1.69M borrowers in China demonstrate it); at maturity it follows the FICO pattern (once repayment is observed, it dominates).
Missing data is never imputed as neutral. A signal with no usable data is excluded and the remaining weights renormalized, with confidence caps — a thin satellite record cannot reach the top band.
This is the part that makes a day-one score responsibly lendable. The tier doesn't just describe the score; it caps what the score may be used for. Rather than burying "we have no repayment data yet" in a footnote, we convert it into an exposure limit in the API response.
We'd rather you knew this before you asked. ATI is a designed model with a live satellite core — it is not a finished production system, and it has never been trained on real defaults.
calibrated: false. No exceptions, no rounding up.The interactive sandbox — move a farmer through ATI-0 to ATI-3 and watch the weights shift into repayment as the relationship matures — is available to partner institutions.