AI Automation Engineer

What Seby is, and why we're hiring

Seby is not advisory. We deploy and we operate. Each customer gets their own always-on agent on its own isolated infrastructure, connected to their email, calendar, chat, meetings, documents and business systems. It drafts replies in each person's own voice, preps every meeting and files every debrief, keeps every contact and deal on one timeline, and takes on the repeatable back-office work that keeps an owner stuck in their own business.

Our customers are owner-run firms: recruiters, agencies, consultancies, design and engineering practices. They pay us, they use it every day, and there are more coming.

We run our own company on the same system. Roadmap, tickets, testing, deploys, alerts and reporting all run through Claude Code and our agents. Two people and a fleet of agents ship what a much larger team would.

We've just signed a large new contract, and demand now outruns build capacity. Our lead engineer is exceptional with Claude and with production systems. We need a second engineer at the same level: someone who can take a customer's mess on a Monday and put a working, verified automation in their hands by Friday.

What you'll build: a real sample

Every one of these is built, or being built, for a customer.

• Onboarding a new customerRender their agent from our one shared template onto its own server, connect Microsoft 365 or Google Workspace, load their contacts on day one, and prove nothing crosses to any other customer.
• Email in their own voiceReply drafts that sound like the person who will send them, check the live calendar before proposing a time, and never go out without the right approval.
• Meeting intelligenceRecordings in from Teams, Fathom or a phone. A prep brief before every meeting, a debrief and actions after, filed against the right contact and company.
• Command centresEvery contact, company, deal and meeting on one clickable, filterable timeline, rebuilt from the customer's real mailboxes rather than a half-updated customer database.
• Messy files into trusted numbersMedia schedules that arrive as scanned PDFs, ad platform exports, timesheets and spreadsheets, turned into clean data with a trail back to the source file and row.
• Advertising automationAn audit that checks an agency's Google Ads accounts against the agency's own house rules, and a campaign builder that shows every change before it goes in, reads it back afterwards, and can roll it back.
• Pricing and project toolsTaking a firm's fee schedules and work breakdowns out of Excel and into a system that prices a tender in minutes.
• Back office and financeInvoices, reconciliations and payments through accounting and banking APIs. Contracts rendered from templates and e-signed.
• Screen recording to specA customer records their screen while talking through a process. An agent turns it into a build brief with frames and timestamps.
• Demos and filmsDemo environments with realistic data, and narrated walkthrough films for prospects.

Sound like your kind of work? See how to apply

How we build

1. Claude Code is the workbench. You'll run several agents in parallel, write the hooks and skills that steer them, and supervise long autonomous runs. If you haven't worked exactly this way yet but you're genuinely fast with AI coding tools, that can work.
2. The builder never marks its own homework. Every claim is checked against the live system by a test, a prover, or a second independent agent. Done means verified on the real surface, not "the code is written".
3. Gates, not good intentions. Rules that matter are enforced in code. A ticket can't be created without testable acceptance criteria. An outbound message can't go without approval. A daily prover checks every live system.
4. Scripts compute, the model reasons, a person judges the exceptions. The model never invents a number. If a figure reaches a customer, it traces back to a source.
5. One template, never forked. An improvement for one customer lands in the shared template and reaches every customer.
6. Isolation by construction. Each customer's data lives in its own space, enforced by the architecture and proven every day.
7. Ticket first, then full ownership. Every piece of work starts as a ticket on our week board. From there it's yours end to end: spec, build, deploy, prove.

How we test: as the customer

Our customers judge us on one thing: does it work when they use it. So every change is tested end to end, as the customer, on the live system, before anyone calls it done. A customer should never be the one who finds the bug.

• Walk the real journey. Sign in as a test user on the customer's own page, click through it, send the message, open the email that lands in the test inbox, and check the record that gets filed. Testing the function underneath is not enough.
• On their devices. Many customers run Windows, and plenty read us on a phone. If it works on your Mac and breaks at phone width, it is not done.
• Test data only. Test accounts and test inboxes, never a real customer's. A test never reaches a real person.
• A standing test for every customer journey. Browser tests and provers run the key paths for every customer on a schedule, so a silent break alerts us before a customer finds it.
• Evidence, not narration. Every delivery comes with its proof: the test run, a screenshot, the live record. Anyone can run it again.

What you'll own

• Customer builds, end to end. From discovery notes to an automation the customer uses every day.
• End-to-end tests for every customer. The browser tests and provers that walk each customer's real journeys on the live system, and the fix when one goes red.
• Integrations. Microsoft Graph (mail, calendar, Teams, OneDrive, SharePoint), Google Workspace, Google Ads, accounting, payroll and HR systems, recording tools and sales systems. OAuth, refresh tokens, rate limits, webhooks.
• Fleet reliability. Scheduled jobs, health checks, and alerts that catch a silent failure before a customer does.
• The template and the evals. So every customer gets smarter when one does.

What good looks like: built, tested end to end, delivered and in use, for every feature. Nothing breaks silently, and no customer finds a bug we could have found first. When you say something is done, anyone can verify it on the live system.

The details

Arrangement

Full-time, senior role, senior expectations.

Location

Remote, with at least 4 hours of overlap with Sydney business hours.

You work with

The founder and our lead engineer, directly.

Stack

Node.js and TypeScript, Python, React 19 and Vite, SQLite with Litestream, Google Cloud VMs, Cloudflare Pages and Workers, Firebase auth, Claude (Claude Code and the API), Whisper, Astro, PostHog, Playwright and the 1Password CLI. Many customers run Windows, so what we ship to them runs on Mac and Windows.

You

• Several years shipping production software, strong in JavaScript or TypeScript and Python.
• You build with Claude Code or a similar agentic tool every day, and can show us something real you built that way.
• You've shipped language model features that people rely on, and you know how they fail: invented numbers, confident wrong answers, drift. You build the check that catches it.
• Real integration scars: OAuth, Microsoft or Google APIs, pagination, rate limits, tokens that expire at 3am.
• At home on a Linux server: SSH, systemd, cron, logs, disk.
• You test like a customer. End-to-end tests are part of the build, not an afterthought: browser automation such as Playwright, test accounts and test inboxes, realistic data. You click through it yourself before you hand it over.
• You finish. You verify on the live system, and you say plainly what isn't done yet.
• Clear written English. Async by default, decisions in writing.
• A massive work ethic. This is a high-output, high-ownership seat in a small company that is growing fast.

Nice to have: recruitment, agency or professional services experience, Microsoft Teams bots, the Google Ads API, accounting or payroll APIs, transcription and ffmpeg, data from messy PDFs and spreadsheets, and agents that run on schedules rather than in a chat window.