Progsu Intelligence Agent

Inspiration

Student orgs run on Drive and Discord. Meeting notes, event playbooks, vendor negotiations, sponsor decks, scattered across hundreds of documents and threads that only the people who wrote them know exist. When someone graduates, that knowledge walks out with them.

I am the Tech Vice President of progsu, Georgia State's largest student tech org. We ran a $20,000-sponsored hackathon with 400 attendees, a Claude Code workshop with 200+ people, and dozens of events across two semesters. At some point we had so much data scattered across so many places that we created an entirely new leadership role: Director of Analytics. One person whose entire job was fetching numbers, tracking data, and answering questions that should have already been answered somewhere. That was the moment I thought: this is not a people problem. This is an infrastructure problem. Why is one person the bottleneck for institutional knowledge that belongs to the whole club?

That is what I built this to solve. Not just for progsu, but for every org running on the same two rails. My goal is not to fix one club. It is to give any student org leader the ability to ask their history a question and get a real answer, without needing a dedicated person to dig it up.

What it does

The progsu Intelligence Agent ingests a full year of real organizational data and makes it queryable and actionable across four modes:

RECALL: Ask what happened. The agent retrieves grounded, cited answers pulled directly from org documents. Every answer shows the source name, date, and a direct Drive link.

ANALYZE: Ask how things have changed. The agent synthesizes across multiple sources and surfaces real numbers and trends. No hallucination. Every claim is tied to a document.

PLAN: Ask the agent to draft something. It retrieves everything relevant, writes a structured planning brief, and creates a real Google Doc in Drive. Not just a text response. An artifact the next person on the team can actually use.

CHAT: Conversational questions handled directly without retrieval.

The practical value: a new officer can ask "what did we learn from Hacklanta" on day one and get a cited, structured answer instead of scheduling a meeting with whoever was there. The Director of Analytics role becomes a query.

How I built it

Stack

• LLM and Embeddings: Gemini 2.5 Flash for generation and reranking, text-embedding-004 for 768-dim vectors
• Agent Orchestration: Google Cloud Agent Builder (Vertex AI Agent Engine)
• Vector and Document Store: MongoDB Atlas

What I learned

Retrieval quality is everything. A RAG system is only as good as the chunks it retrieves. Chunking strategy matters more than model choice. Splitting meeting notes by date header and event docs by heading made a measurable difference in reranker scores versus naive paragraph splitting.

Metadata-aware pre-filtering is underrated. Attaching doc type, event name, and semester to every chunk and using them as MongoDB pre-filters reduced noise dramatically, especially for the ANALYZE query which needs to distinguish Fall 2025 from Spring 2026 data.

What's next for Progsu Intelligence Agent

Every ACM chapter, Google Developer Student Club, IEEE branch, and independent student org faces the exact same problem: institutional knowledge that disappears when leadership graduates. The pipeline is generalizable. Swap the Drive folder ID and the Discord export and the agent works for any org.

Near-term: give incoming exec board members instant access to two years of history before their first meeting. Long-term: a hosted version any student org can connect to their own Drive and Discord in under an hour, with no technical setup required.

Built With

• fastapi
• gemini-2.5-flash
• gemini-api
• gmail-api
• google-calendar-api
• google-cloud-agent-builder
• google-cloud-run
• google-docs-api
• google-drive-api
• mongodb-atlas
• mongodb-atlas-vector-search
• mongodb-mcp-server
• python
• react
• text-embedding-004
• typescript
• vertex-ai
• vite