An AI counsellor, not a lead form
Students researching online degrees in India get chased by call centres the moment they fill a form. This product inverts that: a conversation that earns each question, builds a shortlist matched to the student's goals and budget from verified university data, and never fabricates a fee. Live in production — the product name is withheld here.
Watch a real conversation
This is the live product, recorded unscripted: a student taps one chip, and the counsellor starts — collecting qualification, work status and budget conversationally while the dossier on the right fills itself in.
Recorded on the production site as a tagged test session. The student profile is fictional. Click any screenshot to enlarge it.
Now drive it yourself
A scripted, self-contained simulation of the real flow — tap the reply chips and watch the dossier build, exactly the mechanic a student experiences. Programmes and fees below are fictional.
What we know so far
A journey, not a chat window
The interface is built around a four-stage journey that both the student and the AI can see. A stepper tracks progress on the left; on the right, a live dossier ("What we know so far") shows exactly what the counsellor has understood — every captured fact visible and correctable, nothing extracted silently.
Only after the profile is complete does the product ask for a WhatsApp number — framed honestly as "so I can save what we discuss", verified by OTP, with "Saved, not shared" written right under the field. The gate is a product decision: value first, identity second.
Profile
Qualification, work status, budget and goal — gathered in conversation, shown in the dossier as they land.
Shortlist
Three programme matches from six UGC-approved universities, reasoned against the student's own constraints.
Fees
Full breakdown with EMI options — every number traceable to the university's published fee schedule.
Enrol
Registration and payment happen inside the chat — Razorpay checkout in a message bubble, or the university's own portal.
The surfaces around the chat
The dossier & the gate
Profile captured and visible on the right; the WhatsApp gate appears only once there's a conversation worth saving.
Partner catalogue
Six UGC-approved universities with accreditation badges, programme counts and honest fee ranges.
SEO university pages
A generated, structured page per university — JSON-LD, FAQ schema, canonical fees — feeding organic traffic into the chat.
The three-minute promise
The whole acquisition surface is one search bar: type anything, or tap a chip, and you're in the conversation.
Under the hood
Grounded answers, by construction
Fees, accreditations and programme facts come from versioned data files, never from the model's memory. A pricing guard and per-message hallucination flags catch anything that slips.
Evals as the test suite
LLM behaviour can't be unit-tested, so the repo carries ~60 evaluation scripts — chip quality, pricing accuracy, counsellor tone, PII scrubbing, sycophancy — run against the real prompt stack in CI.
Streaming that survives bad networks
Server-sent events with automatic continuation when responses are cut, retry on model overload, and a fallback LLM provider — the student just sees Sensei typing.
The delayed auth gate
Identity is asked for mid-conversation, after value is delivered — WhatsApp OTP, one field, "Saved, not shared". Gate timing is feature-flagged and A/B tested, not guessed.
Payments in the conversation
Registration fees are paid inside the chat through a Razorpay checkout card — with retry states, gateway abstraction, and per-university configuration for who processes the money.
One codebase, many brands
Partner universities get white-labelled subdomains and an embeddable widget served by middleware rewrites — the same counsellor wearing each institution's colours.