Seminar for groupsNov 2026
A shared AI room where a study group uploads its material, asks one AI together and rehearses for the seminar.
$0.47 → $0.03 cost per question in a room with seven papers: the first after a pause, then follow-ups read from the prompt cache.
The problem
Seminars and group presentations are done together, but students prepare alone, and generic AI chat does not know their material. A group needs one place where everyone sees the same sources, and a way to practise answering hard questions before the real seminar.
What I built
A web app where one person creates a room and the rest of the group joins with a link or a six-character code, no account needed. Inside the room the group shares PDFs, asks a shared AI chat that has read the material and answers with page-level citations, follows a live activity log, and rehearses: individually, or as an AI-led group seminar where each person takes turns answering by voice. Results are visible to the whole group.
The group version took two days to build, on top of a solo practice app I had built first.
How
Next.js 16, TypeScript, Tailwind CSS 4, Supabase (Postgres, Storage, Realtime, anonymous auth), the Anthropic Claude API with the Files API, ElevenLabs for voice, Vercel with a daily cron job.
The browser reads, the server writes.
Row-level security lets a browser read only rooms it is a member of. Every write goes through an API route that checks membership first, with the secret key kept on the server. A script checks that every API route is protected.
Prompt caching by design.
The prompt is ordered stable-first: system prompt, then the room's documents with a one-hour cache breakpoint, then what changes per question. Measured cost per question: about $0.47 for the first question after a pause in a room with seven papers (114,000 tokens), about $0.03 for follow-ups within the hour.
Cost is checked before the call, not after.
Every AI route checks a rolling 24-hour budget for the app and for each room, and logs usage afterwards. Before a chat question the app estimates its cost, including whether the cache is still warm, and refuses the call if it would not fit.
Text instead of PDFs.
Uploaded PDFs are converted to page-marked text before they reach the model. On real files this cut input tokens by 59 %. Scanned PDFs without a text layer fall back to the original file.
The hardest part
Citations the group can trust. Anthropic's built-in citations added 31–46 % more input tokens per question in my measurements, so I had the model write its own compact markers (document, page, quote) instead. The server then checks every quote against the extracted text page by page: a correct marker is kept, a wrong page is corrected, and a quote that cannot be found is removed. This also runs on the live stream, so nobody sees half a marker. In testing, 13 of 13 chat markers and 17 of 17 analysis markers passed, each also checked by hand against the PDF. Clicking a citation opens the PDF at that page with the quote highlighted.
Links
- Demo video above
- Repo available on request