Built by a seller, after a return that did not add up.

We run two businesses on Amazon and Meesho and file our own GST. Like most sellers we used a free tool: upload the marketplace exports, get a JSON, upload it on the portal, done in three minutes. It worked, in the sense that nothing complained.

Then we added up the tables by hand. The liability tables for July carried 3,614.45 rupees of taxable value. The HSN summary in the same file carried 580.58. The entire consumer half of the month was missing from Table 12, and nothing anywhere had said so. The JSON was valid. The portal had accepted it.

The same file also counted two zero-value replacement invoices as cancellations, which understated documents issued and overstated cancellations by two apiece. Small, but it is the kind of small that becomes a question you have to answer in writing three years later.

So we wrote our own. Not a prettier upload form: a pipeline that turns every source into one canonical ledger, builds every table from that single ledger, and then spends most of its effort trying to prove the result wrong. Money is integer paise throughout, so cross-tallies land on zero exactly rather than nearly, because a tolerance is precisely where a real error hides.

Both of those failures are now tests. The actual filed return is loaded as a fixture and the validator is asserted to reject it. That is the standard we would like to be held to.