How we built CNEX-Flow in three months with an AI engineer

People ask how a team this small shipped a platform this big this fast. The honest answer: we built Connekz — our AI engineering teammate — first, then used it to build the rest of CNEX-Flow, with a human reviewing every pull request.

It's not magic, and it isn't "fire your engineers." It's a working pattern you could run tomorrow. Here it is, warts and all.

The working pattern

It's the same loop you'd run with any developer — just faster on the predictable parts:

1. Plan. A human writes a task with clear acceptance criteria.
2. Execute. Connekz reads the criteria and the codebase, writes the tests, then the implementation.
3. Review. It runs the suite and opens a PR. A human reviews and merges.
4. Repeat. It remembers what went wrong last time and doesn't repeat it.

The whole thing hinges on step 3. Connekz opens a pull request — you review it and merge. You're the project manager; it's the developer.

A real task, start to finish

Take a routine one: "Build a CSV contact-import page with column mapping and validation." Clear, well-specified, the kind of work that quietly eats a senior's afternoon.

Connekz pinned the behaviour with tests, wrote the implementation, ran lint + types + the full suite, and opened a PR with the diff. We reviewed it like any other. That's most of CNEX-Flow: the CRM, the boards, the unified call threads, the post-call summaries — shipped this way, reviewed by a human.

// The task, the way Connekz reads it: acceptance criteria first.
export interface Task {
  title: string
  acceptanceCriteria: string[]
  // Tests are written against these BEFORE the implementation.
}

The non-negotiable: you review every PR

Speed without review is just faster mistakes. So the rules never bend:

• Tests first. "Done" means proven, not hopeful.
• Quality-checked before it lands. Work passes a multi-stage review before it reaches your queue.
• You merge. Nothing reaches main without a human saying yes.

That's why the output is worth reviewing instead of being a pile of "almost right." See how the AI engineer works →

What it doesn't do

We're upfront about the edges — it's the honest part, and the useful one:

• Vague tasks get vague results. It pauses and asks rather than guessing, but garbage in still slows things down.
• It won't make product calls. Architecture and trade-offs are human work.
• Net-new, fuzzy R&D is where your seniors shine. Hand Connekz the well-defined work so they can stay on the hard problems.

We keep a running, public list of the limits. See what Connekz can't do →

Why this matters for your shop

You don't need to build a platform to use this. The same loop — write a clear ticket, let Connekz ship the predictable work, review the PR — gives a 5-person shop the throughput of a much larger one, without the hiring. That's the whole pitch, and we're the proof: we run our own company on it.