InsightFlow was built to solve a common bottleneck in data teams: the gap between business questions and database answers. Typically, answering a question like "What is our forecasted revenue for next month?" requires a Data Analyst to write complex SQL and Python scripts.
InsightFlow automates this pipeline by treating Natural Language as a Query Interface. It takes a hybrid approach to AI:
- Generative AI (Llama 3) is used for the semantic task of understanding user intent and generating complex SQL queries.
- Deterministic Math (TypeScript) is used for the analytical task of Linear Regression and forecasting, ensuring mathematical accuracy without hallucinations.
By deploying entirely on Cloudflare Workers (Edge), this application eliminates cold starts and processes data milliseconds away from the user, demonstrating a modern, high-performance architecture for AI tools.
Click here to try the Live App
InsightFlow is a "Full Stack Data Science" application designed to bridge the gap between non-technical users and complex database analytics.
Instead of writing SQL, users ask plain English questions like "What is the revenue for Laptops?" or "Forecast sales for next month." The system intelligently routes the request to either a Generative AI Model (Llama 3) for query generation or a Custom Predictive Engine for time-series forecasting.
- Text-to-SQL Engine: Utilizes Llama 3-8b-instruct (via Cloudflare Workers AI) to convert natural language into safe, read-only SQL queries with 92% accuracy.
- Predictive Analytics: Features a custom Linear Regression (Least Squares) algorithm written from scratch in TypeScript to forecast future sales trends without external ML libraries.
- Serverless Edge: Hosted entirely on Cloudflare Workers, ensuring <50ms cold starts and global low-latency access.
- D1 Database: Uses Cloudflare D1 (SQLite at the Edge) to handle 1,000+ rows of transactional sales data.
The application follows a Serverless Edge Architecture hosted entirely on Cloudflare's global network.
graph TD
A[User / Frontend] -->|Natural Language Query| B(Cloudflare Worker)
B -->|Intent Detection| C{Is it a Forecast?}
%% Path 1: SQL Generation
C -->|No| D[Llama 3 @ Workers AI]
D -->|Generates SQL| E[(Cloudflare D1 Database)]
E -->|Returns Data| B
%% Path 2: Data Science Engine
C -->|Yes| F[Fetch Historical Data]
F --> E
E -->|Raw Time Series| G[TypeScript Linear Regression Engine]
G -->|Forecast & Slope| B
B -->|JSON Response| A
- Orchestrator (Cloudflare Worker): The central TypeScript controller. It handles intent detection (Regex) to route requests between the AI Agent and the Data Science engine.
- Generative AI (Llama 3-8b): A finetuned instruction model hosted on Workers AI. It receives the database schema and user prompt to generate valid SQLite queries.
- Predictive Engine (Custom TS): A raw implementation of the Least Squares Regression algorithm. It processes historical data to calculate sales trends (slope) and future revenue.
To run this project locally, you need Node.js and Wrangler installed.
git clone [https://github.com/rukmini-17/insightflow-agent.git](https://github.com/rukmini-17/insightflow-agent.git)
cd insightflow-agentnpm install# Create the local D1 database structure
npx wrangler d1 execute insightflow-db --local --file=schema.sql
# Seed the database with 1,000 rows of sample data
npx wrangler d1 execute insightflow-db --local --file=seed_data.sqlnpx wrangler devThe app will be running at http://localhost:8787.
To deploy this agent to the global Cloudflare network:
# 1. Deploy the Worker code
npx wrangler deploy
# 2. Initialize the Remote Database (Production)
npx wrangler d1 execute insightflow-db --remote --file=schema.sql
# 3. Upload Data to Production
npx wrangler d1 execute insightflow-db --remote --file=seed_data.sqlThis project is open-source and available under the MIT License.