Capstone: design your own AI helper

Asha is the tech head of Utsav, her college fest in Indore. Last year the help-desk phone rang all day with the same twenty questions. When is the DJ night? Where is the robotics hall? Can I get a refund?

This year Asha wants an AI helper on the fest website. She has a small budget, a team of volunteers, and three weeks. You are going to design it with her, part by part. No code needed: a pen, a page and this course are enough.

So write the first line of your worksheet. Asha’s version: “Answer students’ questions about Utsav’s events, timings, venues, rules and seats, in English or Hinglish. Don’t handle payments, complaints or emergencies.”

Start with the brief, exactly as you learned in “A prompt is a brief to a new intern”. Give a role, context, the task, the format and the tone. Keep it short; every word is sent with every question, so every word costs tokens.

• Role: “You are Utsav’s help-desk assistant for students and visitors.”
• Rules: “Answer only from the fest documents and tool results below. If the answer isn’t there, say you don’t know and point to the coordinator desk near Gate 2.”
• Format: “Reply in three short sentences or fewer. Give times, venues and dates exactly as written.”
• Tone and language: “Friendly and clear. Reply in the language the student used, English or Hinglish.”
• Hand-offs: “For refunds, complaints, medical help or safety issues, give the coordinator or first-aid contact. Do not try to solve them.”

Now the documents. The model has never heard of Utsav, as you saw in “Why AI doesn’t know your data”. So you give it an open book, the idea from “RAG: the open-book exam”.

Asha’s book has five parts: the day-wise schedule, the venue list with directions, competition rules, the refund and entry policy, and a short list of common questions. Only the organisers can edit these files. Students never upload documents into it.

Chunk them by meaning, one event or one rule per chunk, as in “Chunking and vector search”. Then copy the day and date into every schedule chunk. Remember the dog question at the Chai & Chapters café? A chunk that says “Battle of Bands, 7 pm, Main Ground” without “Saturday” can send someone on the wrong day.

Most questions are answered by the documents. One isn’t: “Are seats left for the coding contest?” That number changes every minute, so it needs live data. This is a job for a tool, as in “Tools: giving AI a phone and a calculator”.

Asha’s tool is called check_seats. It takes one event name and returns seats left. It is read-only: it can’t book, cancel or change anything. If a student wants to register, the helper gives them the registration link and the student does it themselves.

Now the bill. Use the meter idea from “Cost and speed: the token meter”. Every question sends the brief, a few chunks and the question in, and gets a short answer out. Here is Asha’s rough estimate for the busiest days.

Two lessons hide in that table. First, the input side costs more than the answers, because the brief and chunks travel with every question. Second, a much bigger model could cost many times more for the same job. A small model with good documents is usually the right first pick for an FAQ helper.

The calculator shows a 30-day month at one price for every token, so divide its bill by 30 to compare with one busy day. Its caching switch treats half of each prompt as a repeat that costs about a tenth. Asha’s fixed brief is closer to a quarter of her prompt, so her real saving would be smaller.

One good demo in the hostel room proves nothing, as you learned in “How to know your AI is good”. Before launch, Asha’s team writes about thirty test questions with the right answer for each, and runs them after every change to the prompt or documents.

Now the safety list, from “Safety, bias, privacy and prompt injection”. Ask what can go wrong, and who catches it. Asha’s team wrote these rules.

• No personal data in the documents: no phone numbers of volunteers or participants, only official desk contacts.
• Chat logs keep the question and answer, never names or phone numbers, and are deleted a month after the fest.
• The one tool is read-only, so even a fooled model can’t book, cancel or pay.
• Emergencies, refunds and complaints always go to people, with contacts shown in the answer.
• The bot is tested in English and Hinglish, and with questions from students of other colleges, so it isn’t only good for insiders.
• Every answer ends with a small “Was this right?” button, and wrong answers are reviewed each evening of the fest.

Step through Asha’s finished worksheet below. For each step, pause and answer it for your own idea before you read hers. Your idea could be a hostel mess bot, a society notice helper or a kirana shop assistant.