Marketing engineering · Growth Platform
Capabilities marketers can run without us.
A working prototype of the two problems in the role: marketing-led landing pages and self-serve audiences. Agents do the patterned work, with evals, monitoring and a defined failure mode before anything reaches a customer.
Read the build-vs-buy memo →1 brief
to a published page
15 checks
before publish
0 tickets
for an audience sync
100%
consent enforced in code
How a landing page gets made
- 01
Brief
Marketer fills a short form. The only free-form input.
- 02
Claude
Writes copy into a typed schema. Picks from approved destinations, categories and quotes.
- 03
Code
Builds URLs and UTMs. Attribution never depends on the model.
- 04
Evals
Up to 15 deterministic checks per market: claims, qualifiers, lengths, audience fit, tracking, alt text.
- 05
Repair
Blocking failures go back to Claude once, with the exact reasons.
- 06
Fallback
Still failing, or API down? Serve the approved template. Never a blank page.
- 07
Learn
A/B test variants, call winners with a stopping rule, and feed what won into the next prompt.
Landing page service
Brief → eval-gated page, with A/B variants
For blog, newsroom, lifecycle and paid teams, in US, UK, French and German markets. Edit inline and the evals re-run as you type. Publishing is blocked until they pass.
Open the studio →Experiment learning loop
A/B results → learnings → better next page
Bayesian readout with a peeking guard. Winning patterns (not exact wording) become memory that Claude uses for the same audience.
Open experiments →Self-serve audiences
Plain English → validated rule → Braze, Meta, Google
Claude proposes, the data contract decides. See size, reach and SQL, then build the destination payload with consent applied.
Open audiences →Design principles
Constrain, then generate
Enums for anything legal or brand would review: destinations, testimonials, categories. The model composes, it doesn't invent.
A defined failure mode
Every path ends in a page that passes blocking evals. The source label tells the marketer which path they got and why.
Contracts in code
Audiences are validated against a typed attribute catalog, compiled to SQL, and consent is enforced at sync for every destination.
Observable by default
Every run logs path, eval results, latency and tokens. Fallback and repair rates are the on-call signals.