We spent months building an AI agent for Terraform. When we did user interviews with SREs and DevOps engineers, their reaction was pretty unanimous: "This looks great but not very useful to me."
This completely changed the trajectory of our startup.
Here is the story of how we accidentally built the wrong product for the wrong people, and how it led us to build Grafos v2: a tool designed to help founders and SWEs survive the transition from MVP to production-grade.
It started as a visualisation tool
During a hackathon a few months ago, we built the foundational blocks of Grafos.ai. Initially, it was just an infrastructure visualisation tool.
As a frontend engineer, I knew nothing about Terraform. I had to go through a hardcore introduction to IaC in just one week. By week two, we had built this:
It was, let's say, "rough" but this had potential. As far as we knew, there were only a couple of decent infra visualisation tools around — Brainboard and Pluralith. The latter wasn't even maintained anymore, and users in their community were practically begging for updates.
But we knew a read-only diagram wasn't the end goal. Not that we understood how to visualise the infrastructure, how hard could it be to build a "Lovable for infra." LLMs know how to code so plugging one in that lets you edit the graph, and translate those edits into Terraform shouldn't be too complicated.
During a second hackathon, we built it. Then we productionised it and shipped Grafos v1. (You can read about the second hackathon here).
We thought we had built a superpower for DevOps but when we talked to actual SREs, their natural skepticism kicked in. There will never be an AI that can claim to be a better SRE than an actual senior engineer. To them our tool was just another abstraction layer to babysit. This forced us to ask who actually needs this.
The answer was staring me in the face. Our ideal user wasn't an SRE. Our ideal user was a founding engineer like me.
Think about the typical founding engineer today. You build a product, you host it on Vercel or Heroku, and you get some traction.
You have a couple of hundred users, and they start complaining about downtime or speed dips. You know you need to migrate to a real GCP or AWS setup. You know of these clouds, but you have no idea what you actually need, let alone how to provision it safely. And you definitely cannot afford to hire a $150k/year infrastructure person yet. Your infra isn't good enough to scale, but you are stuck.
Grafos v2
This is why we threw out our original assumptions and started working on Grafos v2.
From the learnings of our first iteration, we are building a product explicitly to help productionise applications for founders who don't have an SRE. Grafos v1 already knows how to analyze and decide what your application needs. Grafos v2 actually sets up that infrastructure for you based on those requirements.
Because we know the dangers of AI hallucinating cloud infrastructure, we are building v2 with a very strict, opinionated philosophy:
- Highly Deterministic: The system relies on hard logic for as much of the process as possible.
- LLMs in their lane: We only use LLMs to do what they are actually good at. Stuff like reading massive amounts of documentation and interpreting the user's plain-English intent.
- Fail fast to a human: Every step is transparent. If the agent isn't sure, it stops and asks the user, rather than guessing and breaking things.
We aren't trying to replace infrastructure engineers. We are empowering founders to continue their journey, scale their apps, and survive until they reach the point where they can hand it over to a DevOps team.
We're a team of 4 engineers currently deep in the trenches building this. If you are a founder or engineer dreading your infrastructure migration, we are opening up an early alpha for v2 soon to help us test it. Drop your email here to get on the waitlist, or leave a comment below.
Top comments (0)