The AI had to
stay offline.
At Mavenir, I built a tool that answered plain-English questions about telecom inventory. Both the model and the data had to stay on premises. I ran Qwen locally and used LangGraph to translate questions into SQL and route them to the right database.
Read how it worked
I built a Streamlit interface and logged the routing decisions so I could see which database each question reached. I tested queries on a 1,000-row database, where a query took about 20 seconds.
INVENTORY AGENTILLUSTRATIVE QUERY
Which radios are in stock?
SELECT model, quantity
FROM inventory
WHERE type = 'radio'
AND quantity > 0;
READY TO REPLAY
FIG. 01 A sample query through the local workflow
