MigrantBhai

Housing listings, job postings, visa requirements and employer reputation all circulate among Bangladeshi migrant workers — in screenshots, voice notes and group chats. Nothing was indexed, nothing was dated, and nothing survived being forwarded twelve times.

The failure was not a missing database. It was that the people who needed the answer could not tell a current answer from a stale one, and could not read most of the sources even when they found them.¹

The hard problem was never the UI. It was that one worker's "job" is another's "visa category", and a taxonomy that collapses the two produces confidently wrong answers in ten languages at once.

So the schema came first: housing, jobs, a curated FAQ corpus, community threads and the assistant layer, each modelled so that a term translates without losing its meaning. Row Level Security was designed in at schema time rather than bolted on, because the alternative is discovering a leak in production with real people's data behind it.

The obvious build was a chat box on a frontier model. That was rejected: an open-ended generator answering visa questions for people whose livelihood depends on the answer is a liability, not a feature.

Instead the assistant sits over a curated FAQ corpus. It retrieves and phrases; it does not improvise. When the corpus has no answer, it says so.² Every addition to the corpus is reviewed before it is published — the same output-review discipline applied everywhere else in the practice.

The Qatar corridor went live first, with the remaining corridors modelled behind it rather than stubbed. Ten locales route in production, with the content gated per corridor so a worker never reads guidance meant for a different destination.

Built solo — schema design, application, staff tooling and deployment. The interesting constraint of working alone is that every shortcut becomes your own maintenance burden about six weeks later, which turns out to be an excellent forcing function.