Responsible AI
Land records name real people and what they own. These are the principles behind the AI we build for them, and the checks our software puts around it.
Ai GeoLAB builds AI for Bangladesh’s land records. One model reads handwritten khatians (খতিয়ান), the records of rights to land. One turns the old maps of each mouza (মৌজা), the smallest revenue unit, into digital plots. One outlines fields in aerial photos. All three meet in the Khatian Platform, our desktop application for land offices, where most of the principles below are built in.
People approve, not the AI
Our models make proposals. People decide. A reviewer checks every khatian the model reads against the scan and fixes what needs fixing. By default, a reviewer cannot publish their own work. A second person, an approver, has to sign it off, and the platform enforces this four-eyes rule. Mouza maps go through the same review. The AI never publishes anything on its own.
No model is perfect. On khatians from mouzas it had never seen, our reading model scored 91 out of 100 on the index fields, against 16 for the best general-purpose model we tested on the same pages. But 91 is not 100. Some pages are too damaged for anyone to read, and some handwriting is ambiguous even to an experienced clerk. Our Khatian Platform post shows each step.
Show reviewers where to look
We believe a reviewer’s time belongs on the few fields that need a person. The reading model gives a confidence for each field. The fields it is unsure about are highlighted, with a note to check them against the scan. If the same plot (dag) number appears twice in one record, the platform flags it before anyone can approve the page.
If the model cannot read a page at all, the page still lands in the review queue for manual entry. Nothing is lost because the AI failed.
Records stay in the office
The whole platform, including the AI, runs offline on one computer in the office. No scan, record or map has to leave that machine. It works as well in an upazila office with weak internet as in Dhaka. We believe a land office should not have to choose between using AI and keeping its records private. We explain why in AI that stays in the land office.
In the records we show in our posts, owner names are blurred.
Every change can be traced
After a record is published, every change creates a new version and the old one is kept. The platform records who changed what and when, so you can see what a corrected record said before. Each AI reading is tied to the version of the model that produced it.
Tested on what the model has never seen
Every score we publish comes from records, map sheets and areas a model never saw during development. The map model was tested on 22 sheets by draftsmen whose work it had never seen. The field model was tested on unseen areas, then run unchanged over Narayanganj, a district it had never seen.
We report uneven results, not only averages. The map model gets more than nine plots in ten right on the best sheets, and fewer than four in ten on the worst, which are faded and heavily marked. When we found map boundaries sitting slightly beside the ink, a flaw our scores had missed, we fixed it. Across the 22 unseen sheets, the typical distance between our boundaries and those traced by hand fell from 21 centimetres to 11 centimetres on the ground.
We also say where a model falls short. Our field model works well on farmland but poorly in towns, because most of its training labels were rural. We shipped a rural-only version and told its users where to trust it. See How we test our models for more.
No automated decisions about anyone’s rights
Our field model outlines the fields, ponds and homesteads it sees in the 2025 aerial photos. It shows what is on the ground. It does not decide who owns anything. When a reviewer selects a plot, the platform can show the fields the model found on that ground. Where the old mouza map and today’s fields disagree, that is not an error. It is information: it shows where the land has changed since the survey.
None of our models decides who owns a plot, how much of it they own or what it is worth. Those remain decisions for people. Our tools put the record, the map and the ground in front of them.
Bengali first
The platform’s interface is in Bengali, with English one click away. Bengali and English numerals both work in search, and text size can be raised for anyone who needs it.
Keep improving
We are collecting more labels for the field model, including from towns, where it is weakest today. When a model is improved, an office can tell which records the old version read. We believe the people who rely on our models should hear about their limits from us first.
Raising a concern
If you think our models or software have got something wrong, or that we have fallen short of anything on this page, write to [email protected]. Tell us what you saw and where.