Ninaivu

Ninaivu gives a team of support agents one shared memory, and that memory is a graph. Instead of each agent keeping its own chat history, three agents for triage, billing and tech plus a memory curator all coordinate through FalkorDB: who owns an issue, what the customer was promised, and which fix worked last time are all nodes and edges the next agent can read in one query.

The demo follows a fictional customer with a cracked earbuds case. Triage opens the issue and hands it to Billing with a note, Billing starts the refund and records that the customer prefers WhatsApp updates, and a week later, when she asks for an update, Triage answers with the refund status and the WhatsApp preference without asking for the order number again. Every turn shows its tool calls, Cypher queries with timings, and model cost in the interface.

Under the hood it is a stateless FastAPI backend on Vercel, a React front end, and one graph per tenant, so one company's customers can never leak into another's. The load test behind it ran 50 concurrent customers with graph query times under a millisecond, and each customer turn costs about a tenth of a cent on the small model it uses. The project is open source and has no GitHub stars yet.

Honest limits: this is a two-day hackathon demo with fictional customers and synthetic data, not a production helpdesk. The public demo runs on a free database tier that pauses after a day without traffic, so if the graph looks empty the seed script in the repo restores the demo data, and the demo caps free turns per day to keep its model bill bounded.