Early Again: Why I Open Sourced OutcomeOps

Brian Carpio•
OutcomeOpsOpen SourceAI EngineeringContext EngineeringPlatform EngineeringRegulated Industries

This week I open sourced OutcomeOps, the AI engineering platform I’ve spent the last year building, under the Apache 2.0 license. All of it: the PR review, the organizational knowledge base, the code maps, the MCP server, multi-region failover, and the ticket-to-PR code generation. No license key, no paid tier.

The announcement on outcomeops.ai covers what’s in it and how to deploy it. This post is the why, and it starts with a pattern I’ve lived through more than once.

OutcomeOps in action: one question answered across code, ADRs, and the systems an organization already writes in.

I’ve Been Early Before

In 2009 I was running MongoDB in production for telecom carriers while its own logs still said the release was for beta use. In 2012 I built Nibiru at Pearson, a self-service AWS platform that 70 development teams used, before Terraform existed and years before anyone called it platform engineering. In 2014 at Aetna we ran containers on Mesos before Kubernetes was mainstream. At Comcast I made Architecture Decision Records the standard way we recorded why systems were built the way they were, which turned out to be the seed of OutcomeOps.

I’m not listing those to keep score. I’m listing them because being early taught me something that isn’t obvious from the outside: the technology is almost never the hard part. At Aetna we didn’t win by announcing a container strategy. We started with one data analytics team, shipped, and let sprint demos turn skeptics into advocates, and only after that did containers become a corporate mandate. At Comcast our pipeline platform needed an exception to an existing corporate standard before anyone would use it. Being early mostly means spending your time explaining a problem people don’t feel yet.

Early also isn’t the same as right. In 2018 I wrote that Ethereum would become the next generation cloud platform. It didn’t. I try to keep that one in mind whenever I’m confident about where a market is heading.

Where the Market Actually Is

Outside of Amazon, Microsoft, Google, and a short list of AI-native companies, most large organizations are still figuring out how AI engineering fits into their ecosystem. They’ve rolled out a coding assistant. They have an internal chat tool. They have a governance committee writing policy. What most of them don’t have yet is a connection between AI and their own engineering knowledge: the architecture decisions, the standards, the reasons a system is shaped the way it is.

That isn’t a failure. It’s the same place enterprises were with public cloud in 2012 and with containers in 2015. Large organizations adopt new infrastructure once the risk is understood, and for a bank, an insurer, or a health system, that caution is the correct behavior, not a lack of vision. But it means a platform built for the next step arrives before most buyers have a line item for it.

Why Open Source

When a market is still defining the category, most of the work of selling into it is education. A sales process is a poor vehicle for education. Code is a better one. A platform team can read every line, deploy it into a sandbox account, point it at their own repositories and ADRs, and decide for themselves whether the approach works, without a meeting.

The architecture made the decision easier. OutcomeOps always ran entirely inside the customer’s AWS account. Source code, embeddings, chat history, and audit logs never left it, and the models ran on Amazon Bedrock in that account. The only piece that ever talked back to me was licensing. Removing it leaves a platform where everything that runs is in the repository and everything in the repository runs in your account, which is exactly what a security reviewer in a regulated industry wants to hear.

And honestly, I’d rather the ideas get used than sit behind a price tag waiting for the market to catch up.

What Large Organizations Should Push Toward

Whether they ever deploy OutcomeOps or build their own, this is what I think large organizations, especially regulated ones, should be working toward:

  • Standards as an input to AI, not a wiki page. Every AI tool engineers use should be able to retrieve the organization’s decisions and apply them.
  • Review against those standards on every change, with a verdict and a citation, whether a person or a model wrote the code.
  • Your boundary. No source code leaving your environment and no new SaaS vendor in the data path.
  • Auditable by default. Every query and generation logged, with retention you control.
  • Knowledge that compounds. When AI gets something wrong, the fix is a better standard that every future task inherits.

None of that is exotic. It’s the same discipline that made self-service platforms work a decade ago, applied to a new kind of consumer. I wrote recently about how one tightened ADR fixed a governance violation that changing the model never would have. That is the whole thesis in miniature.

What a security team sees: every query audited, every dollar tracked, everything inside your own AWS account.

What Building It Taught Me

Before OutcomeOps, I ran a consumer AI product on my own. Getting a model to follow my standards, and to stay inside safety and regulatory limits, is where I first ran into this problem. It’s also where I learned that you secure an LLM with layers around it, not with a better prompt, which I wrote up in how to build a multi-layer defense system around LLMs.

My first attempts at AI-assisted development on that product produced code I had to refactor about 80% of the time. ADRs were the turning point. Once the model could retrieve the decisions I’d already made, the output started matching them, and the product went from an MVP in 25 days to 90 production Lambda functions by day 120. That approach became OutcomeOps, which I also built mostly alone. I wrote up the operating model behind that.

The lesson I’d hand to any engineering leader is that the model was rarely the constraint. Context was. The organizations that get the most out of AI over the next few years won’t be the ones with the best model contract. They’ll be the ones whose decisions are written down well enough for a machine to follow.

What’s Next

I’ll keep maintaining OutcomeOps: fixing bugs, answering issues, and reviewing pull requests, which the platform also reviews against its own ADRs. The code is at github.com/bcarpio/outcome-ops-ai-assist, and more walkthroughs are on the OutcomeOps YouTube channel.

As for me, I’m taking what I learned into a full-time role, helping organizations make this shift, whether from the inside or alongside them. If the rest of my career is any guide, the market will catch up. I’d like to help it get there.


Brian Carpio has spent 28 years in technology and 14 building enterprise cloud infrastructure on AWS, and ships AI in production daily. He writes about context engineering, governance, and the operating models that make AI-assisted development hold up at scale.