Advising Marketplace

Partner schools required hard data separation, not just tenant columns in a shared database. Students in different markets needed prices calibrated to their region, and availability shown in their own timezone. And the core product moment, a live video conversation, left nothing behind unless the platform captured, summarized, and organized what was said.

On top of that, the money had to be trustworthy: a student paying for a session and the booking never landing, or a booking existing without payment, were both unacceptable failure modes for a marketplace handling real transactions.

Rather than one shared database with tenant columns, each partner school gets its own Supabase project. A master project holds the school registry, advisor profiles, sessions, and payments; at login the platform resolves which database a user belongs to and builds clients for both. Per-school configuration also drives the UI, each school's dashboard shows only the tabs enabled in its registry entry, and a solo-learner mode lets independent students use the platform with no school at all.

Payments confirm the booking, not the other way round. A session is only created after Stripe confirms payment: the checkout call embeds the full session details in Stripe metadata, and the signature-verified webhook creates the session record, stores the amount paid, and credits the advisor's payout ledger. Any processing failure returns an error so Stripe retries, a booking can't silently half-exist.

Session summaries are generated by Claude from call transcripts, which can take a minute or more, too long for a web request. We built a queue on top of the database itself: finishing a session marks it pending, a Vercel cron fires a worker every minute, workers claim jobs atomically, poll until the transcript is ready, call the model through a wrapper with timeouts and exponential-backoff retries, and record per-job status with an admin endpoint for monitoring and manual retry. This replaced an earlier synchronous design that produced client-disconnect errors.

Students browse a directory of advisor profiles, university, major, expertise areas, and ratings computed from session reviews, with search and filtering, plus a Smart Match view that ranks advisors by interest overlap. The booking modal computes the session price live from a base rate, a regional addition resolved from the student's school country or university, the session duration, and a bonus tied to the advisor's accumulated score, with availability converted into the student's own timezone automatically.

Sessions run in the browser on the 100ms video stack: a device-check preview, mute, camera, and screen-share controls, virtual backgrounds, and live captions, with the transcription stream batched into the session record as the call happens. Afterwards, the background pipeline drafts a structured summary, but students don't see it immediately. The advisor reviews the draft, refines it with AI assistance if needed, and explicitly sends it to the student. The platform treats the summary as a deliverable the advisor signs off on, not raw model output.

Continuity comes from session-locked real-time chat enforced by row-level security, file resources and action items attached to each session, and Thoth, a student copilot that answers questions across the summaries and transcripts that have been released to that student, and only those. A Launchpad career board lists internships and jobs with AI guidance per opportunity, counselors get school-wide dashboards and session calendars, and token-gated admin endpoints expose the AI pipeline's queue state with the ability to force or retry jobs.