Beta Lab
Same data, same college, 4 different designs — rebuilt around one idea: answer the question first, then let people dig. Each carries an AI brief that can only cite what the page actually publishes. Real data throughout; things here may still change.
Deep records, a complete FAQ set and a real credited photograph, all present at once. Open any of them in any of the 4 designs.
Indian Institute of Technology Bombay
Vellore Institute of Technology, Vellore
Or open any of 3,925 colleges
The three above are the deepest, but every college in the catalogue renders in every design. Sparse colleges get a shorter page, never filler.
Numbers from instrumented runs on 27 July 2026, not estimates. Where a before-and-after is shown, it is the same page and the same viewport.
Playwright at 1440×900, same page before and after: Coimbatore Institute of Technology. Nothing was deleted — 11 of 12 sections are folded behind a question, one click away.
Time to a complete streamed answer from the three /api/ai/college endpoints. Live, not cached — the cache only serves a last-known-good answer when the model provider is down.
The other 3,389 colleges get a branded monogram, never a stock photograph of somewhere else. Every real photo is Wikimedia Commons or the college's own site, and carries its author and licence on the page.
Proved by accident: on the measured build /brief and /highlights were not yet deployed and returned 404. Every AI surface is additive and client-mounted, so the page kept its hero, snapshot and all chapters.
The model is shown one college's own records and nothing else, and every answer names the sections it drew from — so any claim it makes can be checked against the page it came from.
01
A server-side digest is assembled from this college's own row in our database. The model is given that digest and nothing else — no web search, no training-data recall about the college.
02
The brief and the highlights are requested with a college identifier and nothing else. Asking a question adds your question, and stops there — the page's own text is never uploaded for the model to summarise back to you.
03
Each response opens with a metadata frame listing the sections the answer was drawn from, the model, and when it was generated. A claim you cannot trace to a section is a bug, not a flourish.
04
When a question cannot be answered from this college's records, the reply says so instead of reaching for something plausible. Only the in-scope decision is logged, never the text you typed.
The pages are also built to survive the AI being unavailable: every AI surface mounts on the client after the page is already complete, so a failed request costs you a panel, never the page.