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.
Nautilusthe marketing engineering stack · by Cassien
For performance agencies running twenty or more ad accounts
Know what wins. Make more of it.
Nautilus is the marketing engineering stack we build on your own server. Eight chambers, one loop: every source in one truth, every ad remembered and understood, more of what wins, inside your rules. A person approves every spend.
On your server, under your keysBuilt for Meta AndromedaWorks from Claude Code or Codex
- 1Data Pipelinesevery source your clients use, any API
- 2Data Warehouseall your data and its history, joined
- 3Creative libraryevery ad, described and searchable
- 4Creative DNAthe traits your winners share
- 5Creative loopmore of what wins, paused for approval
- 6Complianceevery variant checked against the rules
- 7Built for Andromedadistinct ideas, each a way in
- 8Your serveryour cloud, your keys, your git
Scroll or press →
01 · The shift
The targeting moved into the creative.
Each time Meta has an ad slot to fill, its Andromeda system narrows tens of millions of candidate ads down to a few thousand, then ranks those for one person.
18.8new creatives a week at the top spend tierMotion 2026
60–370new ads a week for an agency with twenty clientsour math, from Motion
Source: Meta Engineering, Dec 2024
01 · The shift
Every person, a different ad.
The shortlist competes for one slot in one person's feed. Thirty near-copies of one idea give the system one way in; genuinely different ideas give it many. The creative is now the main lever an agency still controls: how much real variety it ships, and whether it knows which variety pays.
Free: the visual patterns behind your clients' longest-running ads, read from the public Meta Ad Library. No account access needed.
02 · Your week
Which number is true?
Meta claims 412 sales. Google claims 389. TikTok and email claim the rest. Add them up and the platforms took credit for 1,181 sales; the store shipped 600. Each one grades its own homework, so someone rebuilds the real figure in a spreadsheet, every week, for every account.
12–24months: when most clients reassess their agencyCallRail, 2026 roadmap for agencies
47%of media-agency professionals are actively considering a career changeState of Media Agency Talent
02 · Your week
Six leaks in a normal week. None is a people problem.
Each one is a missing piece of engineering, and each one has a chamber in Nautilus. Tap one to go there.
- Someone rebuilt the real revenue figure by hand.02Data Warehouse · revenue defined once, in code
- Nobody could say why the winning ads won.04Creative DNA · the traits your winners share
- A brief started from a hunch, not from the data.05Creative loop · briefs written from what won
- Approvals and compliance lived in email threads.06Compliance · rules as code, one approval page
- Client reporting ate whole days.08Ask the warehouse in plain English
- Keys and automations sat on someone's laptop.08Your server · scheduled, logged, yours
03 · The patchwork
Seven tools. Seven logins. Seven versions of the truth.
The usual answer is to rent a tool for every job. Each one is good. None of them knows what the others know, so the joins are still done by hand.
| The job | For example | List price |
|---|---|---|
| Digital asset manager | Brandfolder · Bynder | Vendr medians: $24,999 / $42,500 a year |
| Creative analytics | Motion | $999 / month (Starter) |
| Ad research library | Foreplay | $459 / month (Agency) |
| Data connectors | Funnel | from $300–$600 / month, billed annually |
| Client dashboards | DashThis · Whatagraph | $324 / month · from €999 / month |
| AI creative tool | Creatify | $195 / month (Business) |
| Review & approvals | Filestage · Frame.io | €329 / month · $25 per member |
Our math, at list price, USD tools only: about $1.4K a month for a lean set, about $4.9K with creative analytics and a DAM, about $6.7K with an enterprise DAM. Before seats, overages, and the people who glue it together.
03 · The patchwork
DAM /dam/ · digital asset management
Where a company stores, tags and finds its creative files.
Nautilus includes one. Then it does what a DAM can't.
Every image and video your agency has run sits on your server, described by AI and searchable in plain English. A DAM stops there. Nautilus joins each asset to how it performed, so the library knows not just what is in an ad, but what that ad did. That join has a name: Creative DNA.
04 · A new way
Stop renting a stack. Engineer a system.
Marketing engineering treats growth as an engineering problem, on infrastructure you own. It compounds, because nothing it learns is lost.
- One truthevery source in one data warehouse
- A memory of every addescribed, searchable, yours
- Learning what winsvisual traits joined to results
- Action, inside the rulesa loop with QA and compliance
- Built for the algorithmand running on your server
05 · How Nautilus works
Chamber 01 / 08 · Data Pipelines
Every source your clients run on. If it has an API, it comes in.
Ad platforms, stores, CRM, analytics, email and SMS, payments, call tracking, attribution tools, spreadsheets. Each connector is written as code in your repo for the tools your clients actually use, not picked from a catalogue, and runs on a schedule with a freshness check on every run. Everything lands in one place, so the joins your team does by hand happen by themselves.
WithoutA connector per tool, and the joins done by hand.
With NautilusEvery source in one place, joined, checked on every run.
any API · built on requesthourly where it mattersfreshness check on every run
Chamber 02 / 08 · Data Warehouse
All your data, all its history. Joined and defined once.
Every source lands raw and is kept, long after the platforms stop showing it. Every load is cleaned and tested. The same person in five tools is matched to one customer instead of five, and every order is traced to the ad, the visit and the refund behind it. MER, net CPA, contribution and LTV are defined once, in code, so every person, dashboard and agent reads the same numbers. On a server in your own cloud account.
WithoutFive tools, five versions of every customer, and last year's detail gone.
With NautilusOne customer, one history, one set of numbers your team, your clients and their CFO agree on.
all history, kepttested on every loadone customer, not fivemetrics defined once
How long the platforms keep it: Meta's API returns up to 37 months, Google Ads keeps daily data for 37 months, standard GA4 keeps event-level data for 14 months at most. Meta · Google Ads · GA4
Chamber 03 / 08 · Creative library
Every ad you've ever run, seen and described.
AI watches every image and video and writes down what is in it: who is on screen, what they say, the setting, the text, the second each thing appears. The whole archive becomes searchable the way you would ask a colleague.
WithoutFolders named by date.
With Nautilus“UGC, a face in the first second, talking to camera” → three matches.
images + videosearchable in plain Englishyour server
Chamber 04 / 08 · Creative DNA
What wins, and why.
A DAM knows what is in an ad. Meta knows how it performed. Nautilus joins the two, so you see which visual traits show up in your winners, client by client: a pattern you can test, not a guess. That is the brief for the next round.
WithoutBriefs start from a hunch.
With NautilusBriefs start from what your winners share.
traits × resultsper client
Chamber 05 / 08 · The creative loop
Make more of what wins. Every cycle.
On a cadence set by spend, from weekly to daily, the loop briefs itself from the DNA, makes variants inside the brand rules, runs visual QA and uploads the survivors as paused ads. A person approves. Stop rules follow the limits you set; scaling always takes a click.
WithoutVariants made by hand, checked by eye.
With NautilusVariants made, checked and queued, paused.
briefgenerateQApausedapprovelaunchlearn
42% of consumers say low-quality AI ads would hurt their opinion of the brand. DoubleVerify, 2026
Chamber 06 / 08 · Compliance as code
Regulated clients, without the fear.
Every client gets a rules file. For a fintech: no guaranteed returns, a risk warning on every investment ad, no pressure tactics. Every variant is checked against it before anyone sees it; a break is flagged with the rule it breaks, and it never goes up. Nautilus checks. Your team and your client's compliance lead decide.
WithoutCompliance lives in email threads.
With NautilusRules as code, and a written reason for every flag.
one rules file per clientflag → a person decides
The FCA had 19,766 financial promotions amended or withdrawn in 2024, up 97.5%. FCA, Feb 2025
Chamber 07 / 08 · Built for Andromeda
Meta rewards variety. The loop supplies it.
Thirty near-copies of one ad are still one idea. The loop is designed to send genuinely different concepts, chosen from what your data says wins, so each can find its own people. The results flow straight back into the next brief.
WithoutThirty versions of one idea.
With NautilusDistinct ideas, each a new way in.
distinct conceptsresults → next brief
Chamber 08 / 08 · One system
One system, on your server. It runs while you sleep.
Every chamber runs on one server in your cloud account, behind one API. The routine runs on a schedule: syncs every hour, library and parity checks every day, the creative loop at the cadence spend calls for. Headless AI agents run on the same server, on the same schedule and on your own model keys: they describe new assets, write the briefs, check the variants and draft the reports. No laptop involved, and every run is logged with what it cost.
WithoutAutomations on someone's laptop, and nobody knows what ran overnight.
With NautilusScheduled jobs and headless agents on your server, and a log of every run.
runs on a scheduleheadless agents on your serverevery run logged
Chamber 08 / 08 · How your team works with it
No new app to learn. Your team asks; the system answers.
Nautilus has no screen of its own to log into. People and agents work with it through three doors, all reading the same warehouse:
- Your team Claude Code or Codex. Ask in plain English; the agent runs Nautilus's command line to pull the numbers, run a playbook or write the report.
- Your clients Claude or ChatGPT, through a chat connector (MCP): their own account only, read-only, under your brand, every question visible to their account manager.
- Dashboards One prompt builds one, shared by link with the whole team and live on the warehouse, whoever made it. Your BI tool reads the same warehouse on a read-only login.
WithoutA dashboard per tool, and reports rebuilt by hand.
With NautilusAnswers in the AI tools your team already uses.
06 · Proof, on your data
A 30-day Proof Sprint. Accepted on numbers.
Before you roll anything out, we build Nautilus on five of your client workspaces, each with its own schema and keys, load thirteen months of history and train two of your operators.
- Days 0–3Kickoff, access, a tracking audit (small fixes included)
- Days 3–10Sources live, first Revenue Truth report
- Days 7–14Creative library and Creative DNA
- Days 11–20First creative cycle, up to paused ads
- Days 21–30Your operators run it on their own
06 · Proof, on your data
Accepted when both are true.
✓ The figures match each platform's own reports within ±2% on three key metrics, for at least 28 of the last 30 days.
✓ One of your two operators runs the workflow unaided: a library search, a Revenue Truth report, one loop to paused ads.
If a test fails, we fix it and resubmit it within five business days: that's written into the contract. The fee is credited in full to setup if you sign within 30 days of the sprint ending, and whatever you decide, the data stays in your cloud account.
We run a small number of sprints at a time, so each one gets real attention.
07 · Yours to keep
Not a platform you rent. A system you 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. Our access ends with one command.
- Your brand. Resell it to your clients under your own name. We never contact them.
Not a fit if you want someone else to run your campaigns. We build and maintain the system; your team does the marketing.
What we install
Everything your team needs on day one.
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.
Compliance as code
A rules file per client. Every variant checked against it; every flag carries the rule it breaks. People decide.
Three doors
A command line for Claude Code, Codex and scheduled jobs, a chat connector (MCP) for Claude and ChatGPT, and a read-only SQL login for your BI tools. One API underneath.
Guardrails
Read-only by default. One write path to the ad accounts: every new ad paused, every budget capped, every change approved and logged.
The full toolkit
Seventeen more jobs, on the same warehouse. Every install, every client.
Know what wins
- Creative scorecardsHook, hold, click and convert scores for every ad, with the formulas in your repo instead of a black box.
- Launch tracking and hit rateEvery new ad followed from launch to scale to winner, and the share that make it, per client and per concept.
- Winning combinationsWhich offer, creator, format and landing page win together, across every client's accounts.
- Video drop-offWhere viewers leave each video, second by second, next to what is on screen at that moment.
- Landing-page checkThe ads that earn clicks but lose the sale on the page, with the page to fix first.
Make more of it
- Concept clusters for AndromedaNear-copies grouped into one idea, so you see how many distinct ideas each client really runs, and a duplicate check before launch.
- Angle bankReviews, comments, Reddit and ad libraries turned into the hooks, desires and objections worth testing next.
- Competitor watchWhat each client's competitors launched this week, read from the public Meta Ad Library.
- Creative taxonomyYour naming convention and tags applied to every asset, so any comparison is one question away.
Share it
- Dashboards from one promptAsk for a dashboard and get a link the whole team can open, live on the warehouse, whoever made it.
- Client chat accessYour clients ask Claude or ChatGPT about their own account, under your brand, read-only.
- Branded client portalA login per client, with your name on it.
- Weekly creative digestWhat scaled, what is fading and what to make next, per client, in Slack or by email.
- Growth report and alertsA weekly report per client and an alert the moment a number drifts.
Grow the account
- Lifecycle segmentsRepeat buyers, lapsing customers and churn risk, built in the warehouse and synced to Klaviyo as audiences.
- SEO engineSearch Console queries on page two turned into briefs and drafts, waiting in the CMS for an editor.
- B2B account listsTarget accounts and contacts, built and verified for your B2B clients: data, not sending.
Every playbook runs on the same warehouse and through the same three doors. Care adds a new or improved playbook every month.
Under the hood for your engineers
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
The same person across store, ads, email and CRM matched to one customer, every order traced across tools, and a semantic layer so “revenue” and “CPA” mean one thing to every person and every agent.
Runtime & memory
Scheduled jobs, headless agents and private dashboards on your server, behind your 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.
The full spec
Architecture, security pack and the Proof Sprint terms: read the technical spec.
Also from Cassien · Outbound Engine
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 into their calendar. Run by us; the domains, lists and history stay yours.
- FindCompanies with live buying signals: hiring, funding, moves.
- WriteA personal message per account, in your rep's voice.
- SendDedicated, warmed and monitored inboxes, and LinkedIn.
- ReplyEvery answer read and handled; clear yeses get a time to meet.
- BookInto your rep's calendar, with reminders.
- ReportEvery conversation and meeting, summarised for you.
Also from Cassien · Custom AI engineering
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.
- MapWe sit with the people who do the work and write down where the time and the errors go.
- BuildThe system that does the work, inside the tools you already use, with your keys and your data.
- OperateWe keep it running, measure it and improve it as your business changes.
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. Nautilus is agent-first. Your team works from Claude Code or Codex, which run its command line, and anything visual (a dashboard built from one prompt, a report, an approval page) is published as a private page on your server, behind your single sign-on.
Either works, and your team keeps the one it prefers. Claude Code and Codex both run Nautilus's command line, and the chat connector (MCP) works in Claude and in ChatGPT. Seats and model keys are yours, paid to the vendor directly.
No, and nothing runs on a laptop. The routine (syncs, models, tests, stop rules, the daily digest) is plain code on a schedule on your server. The AI steps run there too, as headless agents on the same schedule and on your own model keys: they describe assets, write briefs, check variants and draft reports. People step in where a decision is needed, and any AI agent you already run can work through the same chat connector.
If it has an API, yes. We don't sell a catalogue: each connector is written as code in your repo for the tools your clients actually use, during setup or as a new source under Care. That covers ad platforms, stores and marketplaces, CRM, analytics, email and SMS, payments, call tracking, attribution tools, reviews and spreadsheets.
Yes, where you want them to. Each client can get a chat connector in Claude or ChatGPT, under your brand: read-only, limited to its own workspace and to the metrics you choose, with every question visible to its account manager. You switch it on per client, where it saves your team the reporting.
It includes one: every image and video is stored on your server, described by AI and searchable in plain English. What a DAM doesn't do, Nautilus does: it joins each asset to how it performed, so you can see which visual traits your winners share.
No. Nautilus is not an attribution model. If you use one, it can be another source in the warehouse. What Nautilus replaces is the patchwork around it: the asset library, the creative analytics, the connectors, the dashboards, the hand-made reports.
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.
Each client gets its own rules file: forbidden claims, required warnings, brand rules. Every variant is checked against it before a person sees it, and a flagged ad never goes up. Nautilus checks; your team and your client's compliance owner decide.
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.
Yes. You can offer it under your own name and at your own price, with a branded portal and a login per client. We stay invisible to your clients and never contact them.
Andromeda is the system Meta uses to narrow tens of millions of candidate ads to a few thousand for each ad slot. It works best with a large pool of genuinely different creatives. That makes real creative variety, and knowing which variety works for you, the main lever you still control.
On a server in your own cloud account. When an AI model describes an asset or makes 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 every 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.
Start here
Know what wins. Keep the system that knows.
Start with a free teardown of the ads working for you, or for your clients, right now. No account access needed.
Prefer email? hello@cassien.ai