InsightFlow

AI Business Intelligence Agent — a production-grade, multi-step AI agent that connects to and takes action across real business apps.

A user asks, in plain language (English, Hindi, or Telugu — text or voice): "Why did sales drop in Hyderabad last week? Tell the team." InsightFlow then:

1. Reads the data from Google Sheets
2. Syncs it and runs deterministic analysis (DuckDB/Pandas — the LLM never does the math itself) with Sarvam-105B reasoning over the verified numbers via LangGraph + LangChain
3. Drafts a report and waits for human approval before doing anything
4. Writes the report to Notion
5. Posts an alert to Slack

Every step is persisted and resumable — a crash mid-run picks back up from its last checkpoint, not from scratch.

Why this project

Most "chat with your data" demos stop at answering a question. InsightFlow is built as a real product: authenticated multi-tenant SaaS (Clerk), billed subscriptions (Razorpay), a durable agent (LangGraph checkpoints on Postgres), and actions across three external systems gated behind explicit human approval — not a notebook, a running service.

Tech stack

Layer	Choice
Agent orchestration	LangGraph + LangChain
LLM	Sarvam-105B (langchain-sarvam)
Backend API	FastAPI (async), SQLAlchemy 2
Database	Neon Postgres (+ pgvector later)
Queue / cache	Redis + Celery
Auth	Clerk
Billing	Razorpay Subscriptions
Frontend	Next.js + TypeScript, installable PWA
Deploy	Render (stage 1) → Coolify on a VPS (stage 2)

Repository layout

backend/     FastAPI + LangGraph backend — see backend/README.md
frontend/    Next.js + React PWA — added in Phase 1 (not yet created)

Status

This project is being built in phases — see the roadmap for the full architecture, phase checklists, and what's done vs. planned. Phase 0 (backend foundations: app structure, Neon wiring, health checks, CI, Docker) is complete; Phase 1 (Clerk auth + multi-tenancy) is next.

Getting started (backend)

cd backend
cp .env.example .env   # fill in your Neon connection strings
uv sync
uv run alembic upgrade head
uv run uvicorn app.main:app --reload   # Linux/macOS/WSL
uv run python -m app.dev_server        # Windows

Full instructions: backend/README.md.

License

Not yet decided / all rights reserved for now.