Orin
Final Submission
Created on 24th April 2026
•
Orin
Final Submission
The problem your project solves
Every meeting in a student-run organization like SPSU produces one shared document - but a Secretary needs documentation gaps, an OC member needs next actions, a CC/VCC needs blockers and urgency signals, and EXCO needs cross-event risk. Nobody needs the same slice. Right now that all gets flattened into one document nobody reads between meetings. The result: committee chairs only notice issues after it's already too late, secretaries spend hours manually reformatting notes into minutes, and important escalation signals get buried in group chats.
How you are solving it
Orin takes a single meeting recording and runs it through a five-agent AI pipeline that produces structurally different outputs for each role simultaneously. The key insight is that we don't just change the prompt - we control what each agent is allowed to see and act on. A transcription agent converts audio to text via Groq Whisper. A meeting agent then extracts every task, owner, deadline, urgency score, and blocker flag from the raw transcript. A routing agent uses that structured state to generate four completely different role-specific briefings. A document agent produces a ready-to-finalize minutes draft. A risk agent surfaces overloaded people, unowned blockers, and deadline risks automatically. All five agents run in parallel and the entire pipeline completes in under 30 seconds. Tasks are saved to Supabase and immediately visible on a live dashboard showing urgency, blockers, and cross-event risk across all active events.
Use of Genspark
We used it to brainstorm on our slides, design and layouts.
Use of Claude
Claude Sonnet 4.6 powers four of the six agents in Orin's pipeline, all running server-side via the Anthropic Messages API. The Meeting Agent reads raw, unstructured transcript text and extracts a fully typed task list - assignees, deadlines, urgency scores (1–5), blocker flags, and status - responding in structured JSON that is validated with Zod before touching the database. The Routing Agent takes that structured meeting state and generates four structurally distinct briefings simultaneously: a documentation-focused brief for the Secretary, an action-focused brief for the OC, a blocker-focused brief for the CC/VCC, and a risk-focused brief for EXCO. The Document Agent converts the same meeting state into formal minutes in paragraph form, ready for the Secretary to review and sign off. The Query Agent answers natural-language questions about live task data with concise, grounded responses. Every Claude call is made server-side only, outputs are Zod-validated before storage, and agents 2–5 run in Promise.all() to keep total latency under 30 seconds.
What is the deployed URL for this project?
https://pushtoprodjava.vercel.app/
Tracks Applied (1)
Hackathon Prizes
Technologies used