AI Data Middleware

Agent-compatible middleware that lets AI systems query live databases through schema introspection, read-only SQL validation, saved connections, and audit logs.

Website: ai-business-analyst-production-8fa6.up.railway.app

Demo: 2-minute system walkthrough

The source application lives in de-10-ai-data-middleware/. This root README is written as the portfolio entrypoint.

What It Does

AI agents are most useful when they can work with real data, but direct database access is risky. This middleware creates a safer control layer:

Natural-language question
        -> live schema introspection
        -> dialect-aware SQL generation
        -> read-only SQL validation
        -> query execution
        -> logged, tabular result

Any agent that understands OpenAI-style function tools can fetch GET /tools/manifest and invoke database actions through POST /tools/invoke.

Why It Matters

Most text-to-SQL demos stop after generating a query. This project keeps the operational parts visible:

• saved database connections instead of one-off credentials
• schema introspection before generation
• read-only validation before execution
• query audit logs
• a browser UI for human review
• a tools manifest for agent integration
• a roadmap toward tenants, source catalogs, object storage, and cost controls

Project Evidence

• Changelog - iteration history from core prototype to control-plane direction
• Roadmap - production-readiness gaps and next steps
• Validation and boundaries - what is proven and what still needs hardening
• End Product Blueprint - SaaS product direction
• Universal Connector Strategy - expansion beyond relational databases

Features

• Natural-language SQL over live schemas
• PostgreSQL, MySQL, and SQLite support in the core demo
• Saved connections with reusable connection_id values
• OpenAI-compatible tools manifest at /tools/manifest
• Tool execution endpoint at /tools/invoke
• SQL validator that blocks destructive and multi-statement patterns
• Schema viewer, SQL runner, Ask AI flow, tools tab, ops status, and history surfaces
• Local query logging plus Supabase-backed control-plane direction
• TokenFirewall integration path for routing, budget, cache, and usage controls

Quick Start

git clone https://github.com/SaneethSunkari/Ai-Business-Analyst.git
cd Ai-Business-Analyst/de-10-ai-data-middleware

cp .env.example .env
# set OPENAI_API_KEY in .env

Start the demo stack:

docker compose up --build

Open:

Surface	URL
Homepage	http://localhost:8000/
Workspace	http://localhost:8000/ui
Swagger docs	http://localhost:8000/docs
ReDoc	http://localhost:8000/redoc

For non-Docker development:

cd backend
pip install -r requirements.txt
env $(cat ../.env | xargs) uvicorn app.main:app --reload --host 0.0.0.0 --port 8000

Demo Database

The local Docker stack includes a PostgreSQL demo database with healthcare-style tables:

• patients
• encounters
• conditions
• medications
• procedures
• observations
• providers
• organizations
• careplans
• allergies
• immunizations
• imaging_studies

Starter questions:

• How many patients are in the database?
• Show provider names and organization names
• List all medications and their total cost
• Top 10 most expensive encounters

API Overview

Area	Endpoints
Auth	POST /auth/signup, POST /auth/login, GET /auth/me, POST /auth/logout
Connections	POST /connections/test, POST /connections/register, GET /connections/, DELETE /connections/{id}
Query	POST /query/ask, POST /query/run
Schema	POST /schema/scan
Agent tools	GET /tools/manifest, POST /tools/invoke
Ops	GET /ops/status

Repository Tour

Ai-Business-Analyst/
|-- README.md
|-- CHANGELOG.md
|-- ROADMAP.md
|-- docs/
|   `-- VALIDATION.md
`-- de-10-ai-data-middleware/
    |-- backend/app/
    |-- backend/tokenfirewall/
    |-- demo_db/
    |-- docs/
    |-- supabase/migrations/
    |-- Dockerfile
    |-- docker-compose.yml
    `-- README.md

Security Posture

This is a portfolio-grade product demo with a serious safety model, not a production security certification. The strongest current controls are read-only SQL validation, connection scoping, explicit tool invocation, and query logging. The roadmap calls out the remaining production work: parser-backed SQL validation per dialect, stronger auth, tenant isolation, secrets handling, policy tests, monitoring, and incident-response docs.

License

MIT. See de-10-ai-data-middleware/LICENSE.