Which ad spend is bringing us customers — and what should we test next?
Follow spend through to outcomes, agree on the next experiment, and give Callie the authority to make the change and check its results.
How work feels today
Somewhere along the way, getting a number became a department. Reports that need other reports. Questions that wait in someone’s inbox.
at Work
Open source On your machines
Meet Callie. She learns how a business works, connects the facts, and stays with the work you hand her. Calliope is going open source. The release is coming. Email Ryan for early access, or for help getting it running on your own systems.
The foundation underneath
Useful decisions need facts you can trust. Callie brings information together, investigates records that disagree, and can collect what’s missing. She works from maintained copies in her own workspace, so that traffic stays off operational systems.
The work, on screen
Follow a real investigation into the data it builds, the work it keeps running, and what becomes possible next.
How she got there · 6 beats
We had advertising results, but couldn’t connect enough of them to the campaigns and settings behind them. The conversation changes the data available, the work that repeats, and what can be answered next.
The gap
The TikTok summary held spend, but most rows couldn’t be connected to the current campaigns and ad groups. Adding a directory of names wouldn’t explain which campaigns produced those results.
The second look
The follow-up asks whether more detailed source tables already exist. Callie looks again and corrects herself: they do. The reporting model was using the less detailed feed. She connects the campaign and ad-group results it had been leaving out.
The broader picture
The request expands to Google and Meta. Callie brings in campaign settings and status, then adds them to the shared reporting data. She checks that the added detail preserves the existing columns and doesn’t multiply the spend totals.
The ongoing work
The enrichment becomes a managed job set to run every twelve hours, calling the connected ad tools through MCP from SQL in Callie’s own database. The collector completes a verified run; the schedule, stored results, and run history stay together.
The next question
Today’s campaign settings don’t tell you what was configured last month. The user raises that limitation. Callie adds a historical capture after each successful enrichment, so future investigations can compare the settings observed over time. That history starts here.
The thing you can use
The expanded data becomes a working cross-platform view. A follow-up brings Salesforce leads and cost per lead into the headline comparison, while campaign-level attribution stays limited to what the data supports. The screen is one result. The next investigation gets the rest.
Definitions, decisions, evidence, and useful methods stay after the conversation. A question can become a working tool. A useful check can become a standing assignment. The next question starts further along.
The everyday work
She notices work worth doing. Let her investigate while you’re elsewhere, then come back to findings and a plan.
The Floor
The Floor keeps what’s underway, what needs you, and what’s finished in view.
The Gallery
The Gallery keeps published work and its sources together, from working views and metrics to scheduled jobs.
Examples of where a conversation can lead, once she’s looking at a business you run.
Follow spend through to outcomes, agree on the next experiment, and give Callie the authority to make the change and check its results.
Trace the gap across products, customers, discounts, and delivery costs. Prepare the follow-up, then keep checking whether the situation improves.
Investigate the accounts, prepare relevant outreach for review, and carry the customer’s reply into the next piece of work.
One workspace for the people who run the business. Callie can keep specialists for sales, leads, and operations — each with its own memory and working history.
You choose the models behind her. American frontier models from OpenAI and SpaceX can take the hard steps; smaller models can take the routine ones. The workspace is yours to run.