AI Admissions Assistant
A conversational assistant that answers admission questions about Amity Lucknow from a curated knowledge base, and declines anything outside that scope.
The problem
Prospective students ask the same few hundred questions about courses, eligibility, fees, hostels and placements — across forms, phone calls and messages. The answers exist in documents; the work is in getting a person to the right one.
What we are building
A retrieval engine over a hand-written knowledge base of admission topics, so the common questions are answered instantly and without a network call.
A scope gate that decides whether a question is actually about the university before anything is sent anywhere. It is enforced in the browser and again on the server.
An optional model-backed fallback for in-scope questions the knowledge base does not cover.
Supporting tools around the core answer path — an eligibility check, a cost estimate and a shareable answer link.
Current status
The most useful thing a project page can do is be specific about what is not true yet.
- A working offline build that answers its knowledge-base topics with no API key and no internet connection.
- A two-layer scope gate, with the server rejecting out-of-scope requests before any upstream call.
- A small Express server that brokers the optional fallback when a key is configured.
- No live deployment serving prospective students.
- No retrieval over official university documents — the knowledge base is written and maintained manually.
- No evaluation set, so answer quality is reviewed by reading rather than measured.
- Not an official university admissions channel, and not a replacement for one.
Technology
Team
Project lead
Position open
Position open
Open roles
Pick the one that sounds like the thing you want to get better at.
- FrontendConversation interface, answer rendering, mobile behaviour
- BackendScope gate hardening, rate limiting, provider abstraction
- AI / LLMMove from hand-written topics to retrieval over real documents
- TestingBuild the question set that tells us whether an answer is right
- UXRefusal states, empty states, and what the assistant says when unsure
What you can contribute
Write questions. The single most useful contribution right now is a list of real questions the assistant gets wrong.
Take one topic area and make its answers precise, sourced and current.
Build the evaluation harness so quality stops being a matter of opinion.
Improve how a refusal reads — that is most of the trust.
Timeline
Define
Complete
Scope fixed to admissions. Everything outside it is refused by design.
Build
Complete
Offline knowledge base and retrieval engine running with no dependencies.
Integrate
In progress
Optional model fallback behind a server-side scope gate.
Evaluate
Next
A labelled question set and an accuracy number we can defend.
Deploy
Later
Only after evaluation. Accuracy before availability.
Want in?
Apply to a role, or start by reading the problem behind it. Neither commits you to anything.