FleetBot
Two builds, one school: a chatbot for 1,500 students, then a staff assistant running on a computer in the front office.
A chatbot for 1,500 students
Every school buries the things students actually need, like the daily schedule, who teaches what, and which clubs exist, inside PDFs nobody ever opens. FleetBot answers those questions in plain English, pulled straight from the school's own documents. I built the whole thing myself and launched it school-wide through the AI Club's Instagram.
The real challenge was making it cheap and reliable enough to run for an entire school at once. My first version burned through $5 every 20 messages, so I rebuilt it from scratch, and the whole public beta ended up costing $2.76 in total across hundreds of messages. I also wrote the rules that stop it from ever guessing an answer or replying to anything unsafe.
An assistant in the front office
The student launch proved it worked, so I pitched the principal on a version for staff. The hard part wasn't the code, it was trust. A school won't run a student's AI unless it's certain the information stays private, so I rebuilt it to keep everything under the school's own control. That's what earned the green light.
The staff version answers questions straight from the school's official documents, and checks all six athletics calendars live to answer “when's the next game?” in about four seconds. I got each answer down to under half a cent and roughly two and a half seconds. A friend hand-coded the staff-facing website and connected it to the system I built.
It went live on a dedicated computer in the school’s main office and handled over 150 real staff questions in its first ten days.
The staff terminal · in the school’s main office“Aamir’s most impressive achievement is FleetBot, a fully functional AI assistant he developed for our school.”
The full engineering record of the student chatbot: three platforms evaluated, the deterministic rebuild that cut cost to cents per message, and the launch numbers.
Read itThe privacy rebuild that won the school over, retrieval tuning that cut a query from $0.45 to under half a cent, and one live integration across six calendars.
Read it