Every tool claims the same sale.
Meta, Google, GA4 and your store each report their own revenue. Someone rebuilds the real figure in a spreadsheet, every week, for every account.
Nautilusthe marketing engineering stack
Know what wins. Make more of it.
Nautilus joins your ad platforms, store and CRM into one truth on your own server, learns why your winning ads win, and makes more of them every week. You approve every spend. The data and the server are yours, and after setup you keep the code.
On your server, under your keysBuilt for Meta AndromedaRunning on our own store first
Connects the tools your growth already runs on
Meta Ads
Google Ads
TikTok Ads
Shopify
WooCommerce
GA4
Search Console
Klaviyo
Stripe
HubSpot
Google Drive
Dropbox
Slack
Claude Code
The real problem
Your marketing isn't short on ideas. It's short on engineering.
Every platform grades its own homework. Creative calls get made on gut feel. AI helps one person on one laptop, then falls over when the whole team needs it. Spend goes up; certainty doesn't.
Nobody knows why the winners win.
Years of ads sit in folders named by date. You can see which ads worked; you can't see which visual choices made them work, so every brief starts from a hunch.
Variety became the targeting.
Meta now matches people to ads from a pool of genuinely different creatives, and near-copies count as one. Most teams can't make real variety at that pace, or tell which variety pays.
AI that stops at the laptop.
Agents run on someone's machine with API keys pasted in. When the laptop closes, the work stops. When someone leaves, nobody is sure what they still have access to.
Marketing engineering
Truth, learning and action. Wired into one system.
Marketing engineering treats growth as an engineering problem: connect every source, define the numbers once, learn from every result and act on it in a loop that never stops. Nautilus is that system: built for you, kept by you.
One number everyone trusts.
Every ad platform, your store, email, payments and CRM in one warehouse, cleaned, matched to one customer and defined once. The figure your team, your clients and your CFO all agree on.
Pipelines · Warehouse · Revenue Truth
Know why winners win.
Every image and video you've ever run, described by AI and joined to its results. You see which visual choices show up in your winners, account by account, and you stop guessing.
Creative library · Creative DNA
Make more of it, every week.
A creative loop briefs itself from what wins, makes new variants inside your brand rules, checks them, and uploads them paused for a person to approve. Built for how Meta's Andromeda picks ads.
Creative loop · Guardrails · Approvals
All of it on your server, in your git, under your keys. We build it, keep it running and train the people who use it.
How Nautilus works
From scattered data to a loop that learns every week.
-
The noise
A dozen tools. A dozen versions of the truth.
Each platform takes credit for the same sale, and every number changes depending on who you ask. Decisions get made on whichever dashboard is open.
-
01 · Pipelines
Every source, piped into one place.
Code-first connectors pull Meta, Google, TikTok, your store, GA4, Klaviyo, Stripe and your CRM on a schedule, hourly where it matters, with a freshness check on every run. Stale data gets flagged before anyone acts on it.
hourly · ads & ordersdaily · GA4 & Search Console13 months of history -
02 · Warehouse
One customer. One order. One number.
Data lands raw, gets cleaned and tested, is matched to one customer instead of five, and becomes metrics you define once: MER, net CPA, contribution. Checked against the platforms every day, on a server you own.
raw → clean → unified → metricsRevenue Truth -
03 · Creative library
Every ad you've ever run, seen and described.
AI watches every image and video and writes down what's in it: the product, the hand holding it, the text on screen, the second it appears. Search the whole library the way you'd ask a colleague.
“product in hand, first 2 seconds” -
04 · Creative DNA
What wins, and why.
Your DAM knows what's in an ad. Meta knows how it performed. Nautilus joins the two, so you can see which visual traits show up in your winners, client by client. That's the brief for the next round.
-
05 · The creative loop
Make more of what wins. Every week.
Every 3–7 days the loop briefs itself from the DNA, generates variants inside your brand rules, runs visual QA and uploads the survivors as paused ads. Your team approves; stop and scale rules follow the limits you set.
briefgenerateQApausedapprovelaunchlearn -
06 · Built for Meta Andromeda
Meta rewards variety. The loop supplies it.
Andromeda is how Meta decides which ad each person sees. It groups near-copies as one ad and rewards genuinely different ideas. The loop feeds it distinct concepts on purpose, chosen from what your data says wins, and the results flow straight back in.
-
07 · One system
One system, on your server. It learns every week.
Your team works through three doors: Claude Code and a command line, a chat connector for colleagues, and read-only SQL for analysts. Jobs and agents keep running when the laptops close, and every run is logged with its cost.
CLI · Claude CodeMCP · chatSQL · analysts -
Why “Nautilus”
Grows without changing shape.
A nautilus builds its shell one chamber at a time. Each chamber is larger than the last, and it carries everything it has built. That's the idea: every loop builds on the last one, nothing you learn is lost, and what it learns stays on your server.
What we install
Everything your team needs on day one.
Revenue Truth
Spend, orders, refunds and returns joined by day, campaign and channel. One net CPA everyone trusts, with an alert when a number drifts.
Creative library + DNA
Every asset described and searchable in plain English, linked to how it performed, with the visual traits behind your winners.
The creative loop
Variants made inside your brand rules, visual QA, upload as paused ads, approval by a person, stop and scale rules you set.
Three doors
A CLI for Claude Code and scheduled jobs, a chat connector for teammates, read-only SQL for analysts. One API underneath.
Always-on runtime
Jobs, agents and private dashboards run on the server around the clock. Every run is logged with what it did and what it cost.
Guardrails
Read-only by default. One write path to the ad accounts: every new ad paused, every budget capped, every change approved and logged.
Warehouse
Postgres with vector search, layered raw → clean → unified → reporting, with metrics defined once and tested on every load. ClickHouse when the volume calls for it.
Pipelines
Code-first connectors in your git repo, incremental and repeatable, with freshness and parity checks so a stale or wrong number gets flagged, not used.
Identity & semantics
Customer matching across store, email, payments and CRM, and a semantic layer so “revenue” and “CPA” mean one thing to every person and every agent.
Three doors
A CLI for Claude Code and cron, a small MCP server for teammates who live in chat, and read-only SQL for analysts, all over one API.
Runtime & memory
Scheduled jobs, headless agents and private dashboards behind single sign-on, with a shared memory your agents keep when people change.
Open by default
Open-source components wherever possible, no markup on the AI tools you use, and a cost log for every call. Nothing to unpick if you leave.
Want the detail? Read the full technical spec — architecture, security pack and the Proof Sprint for agencies.
Ownership
Not a platform you rent. A system you keep.
Yours to keep.
- Your cloud. The server sits in your own account. We don't host your data.
- Your keys. API apps live in your own Business Manager and Google account. No tokens on laptops.
- Your git. Every job, model, playbook and agent skill is code in your repository.
- Your exit. After setup, leave whenever you like: the data is yours, and you keep the code under a perpetual licence, including what your agents have learned. Our access ends with one command.
You're ready if…
- You spend real money on Meta and Google but can't say with confidence which half works.
- Your platforms and your store disagree on revenue, and someone fixes it by hand.
- Your team ships creative volume, but nobody can explain why the winners win.
- You run twenty or more ad accounts and the same report, check and creative cycle repeats for each.
Not a fit if you want someone else to run your campaigns. We build and maintain the system; your team does the marketing.
A week on Nautilus
Six AI agents. One warehouse. One person saying yes.
We built Nautilus on an online jewellery store we own, and that's where it runs first. Here is the week it's built for: six agents working from the same numbers, every creative made around the real product, every spend waiting for one person's yes.
-
Monday 07:30 · Atlas
The morning brief.
Before anyone opens a dashboard, Atlas has read the warehouse: yesterday's spend, net revenue after refunds, MER by channel. One page, numbers everyone trusts, and anything odd flagged first: a tracking gap, a best-seller running low, a new ad beating the average.
-
Tuesday · Iris
New ads, briefed by what won.
Iris briefs the round from the Creative DNA and builds every ad around the store's own product photo, so the product is never redrawn by AI. Visual, brand and IP checks stop the weak ones; the rest land in Meta as paused ads and one tap approves the round.
-
Wednesday · Apex
Budget that follows the data.
Apex reads every ad against the thresholds you set and proposes the moves: scale the winners, pause what stalled, line up genuinely new concepts for Andromeda. Nothing moves until a person approves, and every change stays inside your caps.
-
Thursday · Nova
Customers who come back.
Nova turns the warehouse into segments: bought twice, quiet for sixty days; looked at rings, never ordered; first order this month. She drafts the emails and flows in Klaviyo for review. Sending stays a human decision.
-
Friday · Kairo & Luna
Easy to find. Easy to buy from.
Kairo rewrites the catalog pages Google almost ranks, so search engines and AI assistants can recommend the store; on our own store that work lifted Google clicks 81% in 28 days. Luna answers customers on WhatsApp around the clock. Every result flows back into the warehouse, and Monday's brief starts smarter.
6
AI agents, each with one job, designed to work from one warehouse
0
ads, emails or budget changes that go live without a person's yes
3–7 days
between creative rounds, each briefed by what won the last one
+81%
Google clicks in 28 days from Kairo's catalog work on our own store
An illustrative week. The agents, the warehouse, the guardrails and the ads shown are real; the week shows how they are designed to work together, and its figures and rankings are examples.
How it starts
A 30-day Proof Sprint. On your real data.
Before you roll anything out, we build Nautilus on your own accounts (five client workspaces, for agencies) and your own people run it.
Kickoff, access, tracking.
A 45-minute kickoff and a private checklist of the accesses we need, verified automatically. We audit tracking first; small fixes are included.
Built on your data.
Thirteen months of history from every source, the library indexed, Revenue Truth live and the creative loop running to paused ads. A written update every day.
Your people run it.
Two of your operators search the library, check Revenue Truth and run one loop to paused ads on their own. The sprint is accepted when at least one of them passes that exercise and the figures match each platform's own reports.
Get a free ad teardown
We read the ads running right now for you or a few of your clients in the public Meta Ad Library, tag the ones that have run longest, and send you the visual patterns behind them. No account access needed. The sprint is scoped and quoted in writing after a short call.
Also from Cassien · Outbound Engine
Ask about a seat
The same engineering, pointed at pipeline.
For B2B teams with sales reps. We find companies showing buying signals, write and send in your rep's name across email and LinkedIn, work the replies and book meetings straight into their calendar. Run by us; the domains, lists and history stay yours.
- 01
Find
Companies with live buying signals: hiring, funding, moves.
- 02
Write
A personal message per account, in your rep's voice.
- 03
Send
Dedicated, warmed and monitored inboxes and LinkedIn.
- 04
Reply
Clear yeses get a booking link within minutes.
- 05
Book
Straight into your rep's calendar, with reminders.
- 06
Report
Every meeting in your CRM the moment it's booked.
We run a small number of engines at a time, so each team gets real attention. Every engagement is scoped on a short call and quoted in writing.
Also from Cassien · Custom AI engineering
Request a workflow audit
When the bottleneck isn't marketing.
The same engineering behind Nautilus, aimed at the one workflow that's unique to your business. We map it end to end, build the AI system that does the work, and run it with you.
01Map
We sit with the people who do the work today and write down where the time and the errors go.
02Build
The system that does the work, inside the tools you already use, with your keys and your data.
03Operate
We keep it running, measure it and improve it as your business changes. No prototype handed over and abandoned.
Straight answers
Questions worth asking first.
No. We install it on a server in your own cloud account and maintain it there. You pay once for the setup and monthly for the care that keeps it running; your data never sits with us. If you stop after setup, the data stays yours and you keep the code under a perpetual licence.
No. Your team does the marketing. Nautilus does the heavy lifting up to paused ads, and we keep the system running, update it, repair it and train the people who use it.
Performance agencies running Meta and Google for twenty or more clients, and brands spending seriously on paid social. Smaller than that? Ask anyway; we'll tell you straight whether it pays for itself.
Andromeda is the system Meta uses to decide which ad each person sees. It works best with a large pool of genuinely different creatives and treats near-copies as one ad. That makes creative variety, and knowing which variety works for you, the main lever you still control.
On a server in your own cloud account, behind your own single sign-on. When an AI model describes an asset or generates a variant, that request goes to the model provider set up for your account, and every call is logged with its cost.
You roll it out and we keep it running every month: monitoring, repairs when platforms change their APIs, model swaps, and a new or improved playbook each month. Or you stop: the data stays yours, our code leaves and our access is revoked.
Yes. We're an AI engineering agency: for outbound there's the Outbound Engine, and for anything unique to your business there's custom AI engineering. Tell us what's slowing you down and we'll say straight which fits, or whether none does.
Know what wins. Keep the system that knows.
Start with a free teardown of the ads working for you, or your clients, right now.
Get a free ad teardownPrefer email? [email protected]