Andon monitoring & AI analytics

A factory-floor Andon monitoring platform built for Mahindra Aeronautics — React and Node end to end, with role-based access for everyone from operators to plant management, and an AI analytics chatbot that lets non-technical staff query production data in plain language.

The problem

Andon systems surface line stoppages and alerts in real time, but the data they accumulate is only useful if people can interrogate it. Management wanted answers like "which station had the most downtime last week?" without waiting on someone to write SQL — and without giving an LLM the keys to a production database.

The design

Monitoring core. A React/Node platform streaming live Andon state to the floor and to dashboards, with multi-level RBAC so operators, supervisors, and management each see exactly the scope they're responsible for.

Text-to-SQL chatbot, treated as a security problem. The analytics chatbot is powered by Gemma translating natural-language questions into SQL. An LLM that writes queries against your production database is an injection surface, so it is sandboxed in three layers:

1. Prompt design constrains the model to the analytics schema and read-only intent.
2. Regex filtering validates the generated SQL before execution — anything that isn't a plain SELECT against permitted tables is rejected.
3. A read-only, scoped database role is the backstop: even a query that slipped past both layers physically cannot write, and cannot see tables outside its scope.

Dynamic JSON→charts engine. Query results come back as structured JSON and a rendering engine picks and configures the right visualization at runtime — so the chatbot can answer questions with charts nobody explicitly built a dashboard for.

Results

• Multi-level RBAC across operator, supervisor, and management roles
• Natural-language analytics with defense-in-depth around the LLM
• Charts generated dynamically from any query result shape