Shwetal Mehta

SD-WAN "Rosetta Stone": one engineer, three vendors

SD-WAN Rosetta Stone An engineer's common request goes through the Rosetta Stone: a prebuilt library handles about 80% of work and GPT the rest, producing commands for Versa, Viptela and Meraki that the engineer reviews and applies; tier 1–2 resolution time fell 50%. Tier 1–2 resolution time −50% Common task: "Add a site VPN" Rosetta Stone Prebuilt command library ~80% of work GPT for the rest ~20% Versa commands Viptela commands Meraki commands Engineer reviews and applies
SD-WAN Rosetta Stone An engineer's common request goes through the Rosetta Stone: a prebuilt library handles about 80% of work and GPT the rest, producing commands for Versa, Viptela and Meraki that the engineer reviews and applies; tier 1–2 resolution time fell 50%. Tier 1–2 resolution time −50% Common task: "Add a site VPN" Rosetta Stone Prebuilt command library ~80% of work GPT for the rest ~20% Versa commands Viptela commands Meraki commands Engineer reviews and applies
An engineer describes the task once; the library or GPT returns each vendor's commands, and the engineer applies them.
  • −50%tier 1–2 resolution time
  • 3vendors, one procedure
  • ~80%handled by a prebuilt library

Why I built it

We supported three SD-WAN vendors (Versa, Viptela and Meraki), and each spoke its own configuration language. Tickets waited for the one specialist who knew the right syntax.

What it does

  • An engineer describes the task in common terms.
  • A prebuilt library returns the exact vendor-specific commands and configurations for about 80% of repeatable work.
  • For the rest, OpenAI GPT generates vendor-specific commands, which the engineer reviews and applies.

How I built it

  • The idea and the sponsorship were mine, and I built the prototype myself in late 2024. My operations team built the production version.
  • A deterministic library first; the model only for the long tail. It's the same principle as the pricing engine.

What I learned

  • Use code for the common case and AI for the exceptions.
  • Keep a human applying changes to customer networks.
  • Give a tool a name people remember.