About the roleWe run an internal operations platform that a mid-sized manufacturer depends on daily — production scheduling, warehouse transfers, inventory counts, PO tracking, packout, and shipping. About fifty people use it, including floor staff on scan guns. It connects to an ERP and a Postgres database, and it changes almost daily.

Most of the code is written by AI coding agents. That part works. The bottleneck is the human loop around them: intake, triage, scoping, prompt-writing, reviewing what the agent produced, verifying it works, and keeping the documents agents read from current.

That loop is this job. You're not the developer. You're the person who turns "this button's broken" into a scoped, verified, shipped fix — agents do the building, your judgment makes sure it's right.

And you don't do it alone. You'll use AI to manage the AI — to draft prompts, explain a diff you can't read, summarize what changed, check an agent's work against what was actually asked for, and write up documentation. Knowing how to put a model to work on your own behalf is the core skill here.

What you'll doIntake and triage. Turn vague bug reports from across the business into tickets with a repro path and a sample record. Reproduce before diagnosing. Sort ruthlessly: wrong number for a user today, or backlog?

Ship fixes end to end. Write the agent prompt with an explicit scope contract — what's in, what's out, when to stop and ask. Check the result against that contract, then drive it through PR, deploy and live verification.

Translate. For every change: what it does, what it means for users, what needs deciding, what breaks if it's wrong. Use AI to help you write it — that's expected, not cheating.

Support larger builds. Break direction into phases that ship and reverse independently. Mock up interfaces before UI work starts. Prepare database migrations as step-by-step handoffs for someone else to run.

Operate the agent fleet. Keep agent environments clean and their context documents current. Track which agent runs where, with what credentials — agent config is a permission surface, and widening it needs approval.

Keep documentation honest. Update project memory and the release log after anything ships or pauses, including what was not verified.

What you need to know about AIYou don't need to be an AI expert. You need to be someone who genuinely works this way already, and who uses AI to get more done than they could alone.

What we're looking for





- You've used an AI assistant on real work — coding tools like Claude Code, Cursor or Copilot, or general assistants you've pushed hard on real projects. Formal job experience is a plus, not a requirement.
- You direct a model; you don't just ask it questions. You give it context, set boundaries, and steer it when it drifts. An unbounded prompt produces unbounded results, and you've figured that out.

- You check the work. You know models sound confident when they're wrong — invented file names, plausible summaries of things that didn't happen — so you verify before you pass anything along.
- You use AI on your own tasks, not just the build. Drafting a prompt, explaining code you can't read, turning a messy user complaint into a clear ticket, writing documentation. If you're not already doing this, this role will be a grind.

- You're comfortable being the non-expert in the room and using AI to close the gap quickly, without pretending you know more than you do.

Helpful, not requiredRunning more than one agent at once and keeping them from colliding · working directly with a model API · MCP servers or connectors · retrieval and context engineering · testing whether a prompt change actually improved anything · supporting an AI assistant used by non-technical people. Not needed at allWriting production code — you need to read a diff and explain it plainly, and AI can help you do that.

No machine learning or data science background. No CS degree.

Also helpful

- Basic Git comfort — branches, pull requests, merges. Enough sense of what goes wrong to stop and ask rather than improvise. We'll teach you our specific setup.
- SQL as a reader — or the willingness to get there. You won't run writes or migrations; that authority sits elsewhere, deliberately.

- Writing that respects the reader. Bullets. Recommendation first. One decision per ask. Long paragraphs don't get read here.
- Willingness to say "I don't know." A confident wrong answer costs more than an admitted gap. This matters more than any technical item on this page.

- Care about the people on the other end. A wrong number sends someone to the wrong bin or ships the wrong pallet.

Nice to have:




- Manufacturing, warehouse, ERP or WMS exposure
- Node/Express, vanilla JS, Postgres, REST APIs, AWS, GitHub Actions, Microsoft 365, Airtable, QuickBooks
- Internal documentation people actually follow · supporting non-technical users.

How we workWe move fast because the guardrails are clear. Most exist because something went wrong once.
- The gate is the plan, not the code. State the plan in plain English, get approval, then execute autonomously inside that scope.

If the plan turns out wrong mid-build, stop and re-ask.

- Fail loud, never fake. A visible failure beats a silent one. Placeholder data to make something look fine is never acceptable.
- Never assume — verify.

Not that a migration ran, not that a document matches reality, not that an agent did what it said.

- Product decisions aren't yours. What a tool does, the business rules, the interface and the wording belong to the product owner. You bring recommendations, not mandates.
- Some things are off limits to this role — running migrations, changing auth or permissions, handling secrets, and a short list of irreversible operations.

Not a trust problem; it's how the system stays recoverable.

- Scope creep is the failure mode we watch for most. Helpfulness beyond what was approved is still unapproved work.

First 90 daysDays 1–14: Full context system read.

Shadowing intake. Three low-risk fixes shipped through the complete review-and-verify loop. Days 15–45: You own daily triage. Prompts written unassisted, with only the plan reviewed. Verification write-ups and screenshot sign-off are yours.

Days 46–90: Backlog owned end to end. One medium build phase planned. Agent hygiene and config inventory on a standing cadence you maintain.

Measured by: bugs closed per week, and fixes that don't come back · time from report to reproducible ticket · zero unapproved scope · zero guardrail violations · verification gaps disclosed, never hidden · documentation current enough that the next agent session starts from truth.

How to applyEmail a short note to [email protected] covering:

1. Something real you got done with AI — what it was, how you set the model up to succeed, and what you had to correct.
2. A time you caught AI being confidently wrong. What tipped you off.
3. How you use AI in your own day — the unglamorous parts, not the demos.

Short beats long. We'll follow up with a practical exercise using a real example from our backlog.

📌 Bilingual AI Development Assistant (Colombia)
🏢 Merit Manufacturing
📍 Colombia