AI Automation Engineer

The short version

MnemioAI gives every employee an AI agent that works in the apps they already use and learns how they work. Your job is to make those agents genuinely useful inside real businesses: sit with a team, find the work that costs them the most, and replace it with agents that are accurate enough to trust.

You'll be half engineer, half business analyst. The engineering gets the agent working. The business sense decides which work is worth automating first.

The bar: better than AI, using AI

Give a capable model a task with no context and you get a plausible, average answer. Give the same model the right data, the business's rules, examples of good and bad output, tools that check its work, and someone who knows where it tends to fail, and the result is a different class of work.

That difference is the job. We judge the work on one question: is what the agent produces with your setup clearly better than what the customer would get by asking the model directly? Better means more accurate, closer to how this particular business works, and finished enough that a person approves it instead of redoing it.

In practice, that means being comfortable with the unglamorous parts: reading the data before writing a prompt, building evaluations, testing the edge cases, and saying "the model can't do this reliably yet" when that's true.

What you'll do

Map how work actually happens in a customer's finance, operations or sales team, including the steps nobody wrote down.

Rank processes by hours lost, errors and money at risk, against the effort to automate them. Start with the ones that pay back in days.

Build the automations on MnemioAI: connect tools through APIs and MCP, write the agent's instructions and skills, add approval steps so people stay in control, and handle the messy cases.

Measure the outcome (hours saved, errors caught, cash recovered) and report it in language a business owner understands.

Bring patterns back to the product team, so the fifth customer gets in a day what the first one took a week to get.

What a good week looks like

On Monday you sit with a finance manager who spends nine hours a week matching supplier invoices to purchase orders and delivery notes. By Thursday an agent does the matching overnight and flags the six that don't line up, with the reason for each. She clears them in twenty minutes. On Friday you write down what you learned, so the next customer with the same problem is live in a day.

Problems you'll work on

Matching invoices against purchase orders and delivery logs, then preparing bills in QuickBooks, Xero or Tally for approval

Reconciling marketplace payouts against orders, fees and refunds, and drafting claims for money owed

Turning CRM, ad-platform and finance data into a weekly management report, with the commentary drafted

Chasing overdue invoices in each client's channel and tone, and knowing when to hand over to a person

Answering employees' questions from the company's own documents and Slack history, with sources

What we're looking for

2–5 years building software, automation or data workflows in Python or TypeScript, with real APIs, webhooks and messy data

Hands-on work with large language models: prompting, tool use or agents, evaluation, and a clear sense of where they fail

Evidence that you get better results from AI than the default: a project, a before-and-after, an evaluation set, anything you can show us

Business sense: you can sit with a finance or operations lead, understand the process quickly and tell which problem to solve first

A bias for return on effort: a two-day automation that saves ten hours a week beats a three-month project that might

Clear writing and speaking in English, and patience explaining technical trade-offs to non-technical people

Nice to have

Building MCP servers or integrations for business software

You'll probably enjoy this if

You've always been more interested in why a process is broken than in which tool fixes it

You check a model's output before you trust it, and enjoy working out why it was wrong

You like seeing real people use your work in the same week you built it

Apply here directly: [hidden]