← Back to projects

Project

FinOps Multi-Agent Assistant (in progress)

Multi-agent assistant over finance workflows (in progress): a router classifies intent and dispatches to specialized agents, with typed request/response contracts, trace IDs and per-step latency events. The Angular UI and the FastAPI routing/observability skeleton are deployed; the LLM, RAG and tool layers land next — I'm building tracing and cost telemetry first so agent answers ship measurable from day one.

In progress Own product — in development Engineer (Architecture, Frontend, Backend Scaffold) Public

Own project built as an architecture study, not a production system. The router, API contracts and observability events are implemented; agent responses are still stubs pending LLM, RAG and tool integration.

Live / Demo GitHub

Stack

PythonFastAPIPydanticAngularTypeScriptSCSSFirebase HostingGCP

Impact

  • Delivered a deployable Angular landing + demo chat for an enterprise-style multi-agent assistant.
  • Designed explicit agent roles (Router/Task/RAG/Tool) to enable safe orchestration and future backend integration.
  • Established a clean MVP path to production deployment (Firebase Hosting + Cloud Run backend planned).

What I did

  • Designed the agent topology — Router, Finance Analyst, RAG Policy and Tool agents — with explicit orchestration instead of hidden chains.
  • Implemented the FastAPI surface with typed Pydantic contracts, a stateless /v1/chat endpoint and health checks.
  • Every response carries a trace ID and an ordered event list with per-step latency, so the routing path is observable from the UI.
  • Layered the backend by concern (domain / services / infra / api) so each agent and its infrastructure can be replaced independently.
  • Built and deployed the Angular UI to Firebase Hosting, including a demo chat that renders the agent trace.

Architecture

  • User → Router Agent → task agents → tool layer → aggregated response
  • FastAPI app split into api/v1 (routes), domain (schemas), services (agents) and core (config, logging)
  • Stateless chat endpoint: session ID in, answer plus events/steps/tool_calls/citations out
  • Intent routing is deterministic today (keyword classification) and is the seam where the LLM classifier plugs in
  • Planned: Chroma vector store for the RAG agent, a bounded tool-execution layer, Docker image and Cloud Run deployment

Tags

AI AgentsA2ARAGLLMsFastAPIFinOps