Fraqt

An international student office runs on deadlines that carry legal consequence. A date read wrong from a form becomes a status problem for the student, well after anyone could have caught it. The offices doing this work are small, and the questions arriving at them are repetitive, specific, and expensive to get wrong.

Fraqt answers those questions from a corpus of federal guidance, refuses when the corpus does not cover them, and drafts officer replies that a person edits and sends. Built by one person in six weeks from an empty repository to production.

They sign up. The first thing it asks for is their I-20. Upload from a laptop or photograph it with a phone.

Then a screen with the document image on one side and the extracted fields on the other. SEVIS ID, program start and end, school, country. They read across, fix anything wrong, and confirm. Nothing the system computes touches a field they have not confirmed.

Then the dashboard. Their name, their school, their status. Underneath, one to three deadline cards. Only the deadlines that apply to them, each with a why that opens the passage it came from.

They type a question. “Do I need to file Form 8843 if I had no income?” The answer streams in, plain language, with a sources list underneath showing each document, its date, and a link. If the corpus does not cover it, they get a calm card instead. This depends on details we cannot see. Your DSO can help.

That refusal is the product working, not failing.

They open the queue on Monday morning. Student threads, open and unassigned. They click one. A student asking about Form 8843.

Below the message is a button. Draft a reply. They click it, and a reply appears in the box they already use. Cited, grounded, plain. They read it, fix a sentence that sounds off, and hit send. The student gets it as a normal message from their office.

The officer never sees the AI. They see a draft they can edit and send.

Quiet. No urgency colours, no chatbot personality. Closer to a well made government form than a consumer app. That is deliberate. The people using it are already scared.

Textract reads an I-20 and gives back field names that change with the scan, the angle, the school’s printer. The same date field comes back under three different labels. For a while nothing mapped cleanly and the dashboard showed garbage.

The fix was not clever. It was accepting that extraction is unreliable and putting a human confirmation step in front of every computed value. That decision came out of the failure, and it is the reason confirm-before-use is an architectural rule here rather than a preference.

Analytics events were emitted inside the database transaction that recorded an officer’s decision. A telemetry failure could therefore roll back a real approval. An officer would have clicked approve, seen it succeed, and had it silently undone because a metric could not be written.

It was found by reading code. The tests passed. They passed because they exercised the approval path and the analytics path separately, and the fault only exists where the two share a transaction boundary.

The fix was ordering. The emit now sits after the commit, and its failure is logged and swallowed. The work goes through and the number is lost.

There was a second one of the same shape. Every draft analytics event was silently dropped in production because the development tenant identifier failed a strict validator, and the tests passed because they used a hand-written value instead of the real constant.

Both are stories about tests passing while the system was wrong. Green tests do not mean the page works.