Skip to content
~/work/ramanujan

Ramanujan

Agentic assistant · Microsoft Teams

An AI analytics copilot: business teams ask questions in plain English inside Microsoft Teams and get answers from enterprise data in seconds, with no SQL and no dashboards.

What it does

  • 01A Gemini ReAct agent turns natural-language questions into governed BigQuery SQL.
  • 02RAG over a Qdrant-backed KPI glossary with sentence-transformer embeddings, so answers match the business definition of each metric.
  • 03A production FastAPI service on Vertex AI with hexagonal architecture.
  • 04Safety layers: a SQL policy engine with row-level access control, and grounding guardrails against hallucinated answers.
Stack
Python · FastAPI · Gemini · Vertex AI · BigQuery · Qdrant · Sentence-Transformers · Microsoft Graph API · Docker
Domain
AI / ML · APIs
Ships as
Agentic assistant · Microsoft Teams

Architecture

askKPI definitionsSQLgoverned querygrounded answerMicrosoft TeamsQuestions in plain EnglishL1FastAPI serviceVertex AI, hexagonalL2Gemini ReAct agentQuestion to SQLL3KPI glossary (RAG)Qdrant + embeddingsL4SQL policy engineRow-level accessL3BigQueryEnterprise dataL4

Simplified from the project description.