Skip to content

Selected work · AI · Full-stack · 1 min read

AURA: a five-agent finance advisor

Google Cloud Agentic AI Day finalist, top 3.5% of 57,000+ participants, as team lead.

Visit live site GitHub Watch demo

TL;DR

  • Team lead; Google Cloud Agentic AI Day finalist, top 3.5% of 57,000+ participants.
  • An orchestrator runs 2, 3 or 5 agents in parallel with 20 s timeouts and streams answers over Socket.IO; an MCP client exposes 6 finance tools.
  • The live build uses keyword retrieval; next step is vector retrieval with an eval set.
Role
Team lead
Team
Hackathon team
Timeline
2025

How it fits together

Request

  1. Rate complexity
  2. Run 2, 3 or 5 agents in parallel
  3. MCP finance tools
  4. Stream over Socket.IO
Simplified view of a request.
AURA: the live site's home page
The live site, as of 3 Nov 2026.

How it works

An orchestrator rates each query's complexity and runs 2, 3 or 5 agents in parallel, each with a 20-second timeout, streaming the answer over Socket.IO.

An MCP JSON-RPC client exposes 6 finance tools: net worth, bank transactions, mutual-fund and stock transactions, credit report and EPF.

Google OAuth, Helmet and rate limiting; deployed on Azure App Service.

Results

  • 3–8 s

    answers

    On the hackathon dataset, not a production SLA

  • Top 3.5%

    of 57,000+ participants

    Google Cloud Agentic AI Day finalist (2,000 finalists)

What I'd do differently

  • The deployed version uses keyword retrieval over finance guides; embedding retrieval and MongoDB were disabled in the live build.
  • Next step: re-enable vector retrieval, with an eval set to prove it helps.

Stack

  • Node.js
  • Express
  • Azure OpenAI
  • MCP
  • Socket.IO
  • Google OAuth
  • Azure App Service