Case study · f. 02r
Study Material Assistant
- Client
- Bisma Ghani
- Year
- 2026
- Timeline
- Working version in 3 days
- Stack
- Python, RAG, Citations, PDF parsing
- First working version3 days
- Answers without a sourceRefused
- Source shownEvery answer
The problem
Generic chatbots answered confidently from the internet, not from the course. For exam prep that's worse than useless: a wrong answer that sounds right.
The client needed to trust the assistant, which meant seeing where each answer came from and being able to check it.
The approach
- 01
Scope first, in plain English
Before writing code we agreed on what "correct" meant: answer only from the uploaded material and show the source, otherwise say "not in your notes".
- 02
Citations as a feature, not a footnote
Every answer links back to the exact document and page, so the student can open the original and verify it in seconds.
- 03
Flagging risk early
One requested feature would have mixed outside knowledge into answers and quietly broken accuracy. We flagged it on day one and designed around it instead.
The outcome
A working version was delivered in three days. Answers the material can't support are declined, not invented.
Delivered a working version in 3 days. He flagged risks early, explained technical trade-offs in plain English, and even told me when one of my ideas would break the accuracy. He was right. Exactly what you want in an AI developer.