Milestone 2 - What broke, what I fixed, and what 10 minutes of my time produced.
If you missed milestone 1 - I built an agentic marketing system
in 3 days to run Crevaxo's marketing while I work a full time
physically demanding job. That post is here: https://www.indiehackers.com/post/milestone-1-i-built-an-ai-workforce-to-run-my-saas-in-3-days-heres-exactly-what-i-built-and-why-66d9ecfd98
This is the honest milestone 2 update.
What shipped
2 full campaigns. 12 posts generated, approved, and scheduled
across all content pillars. Total time I spent: roughly
10 minutes across both campaigns - brief drafted on my phone
between jobs, approvals tapped through on Slack when they came in.
Nothing posted without my explicit sign off.
What I fixed
Scheduling intelligence was broken in the wrong direction.
The research agent was correctly identifying optimal posting
windows - Wednesday 9am AEST for Crevaxo's audience - but the
scheduler was treating that as a loose suggestion rather than
instruction. Fixed by making the research agent return
structured JSON that the scheduler consumes directly.
The data it generates now actually determines what gets scheduled.
Flat rejection was a dead end. Original approval flow was binary -
approve or reject. A rejected post taught the system nothing.
Added a three option flow - approve, revise with reason, or discard.
Revision reasons are passed back to the content agent as explicit
instruction on the next attempt. The decisions log is building
into a real training signal over time.
Memory layer wasn't being used. Fixed the content agent so it
queries 21 days of post history before generating anything -
pillar rotation enforced, angle variation enforced, performance
signals from Zernio analytics feeding in as context. The
compounding effect is slow but it's there from day one.
Silent failures with video content. When a weekly plan
included a reel the image generation model was being called
on a video asset type and failing without surfacing anything.
The worst kind of failure - you only notice it's missing.
Fixed with a reel flag - production brief and a Slack
notification surface instead of a broken generation attempt.
Video generation APIs are mature enough to solve this properly
(fal.ai gives single endpoint access to Seedance, Kling, and Veo).
That's the next infrastructure decision.
Architecture decision - moved from PostFast to Zernio for
social scheduling. PostFast works but the pricing model
doesn't scale well with user accounts. Zernio charges per
connected social account with a generous free tier and
per-account pricing that gets cheaper at volume. Zero
scheduling cost at current usage.
Honest numbers
1 waitlist signup.
2.8k views on the milestone 1 Reddit post.
12 posts shipped through the system across 2 campaigns.
0 usage of the free text chat option on the landing page -
probably a visibility problem, working on it.
What's next
The onboarding agent is the priority. Right now every
new user needs manual setup - I write the brand context,
configure the workspace, run the first campaign manually.
That doesn't scale past 3 or 4 people. The target is
URL input to first approved post in under 10 minutes
with no manual intervention. It's the most complex build
in the system and the one that unlocks external users.
Milestone 3 will have real external user data.
That's the post worth waiting for.
Building this in public. The good, the bad, and the honest.
Nothing is off limits.
This is getting more interesting because the product is moving beyond “AI content generation” into an approval-driven marketing operating system. The revise-with-reason flow is the strongest part here. That turns human feedback into training signal instead of treating approval as a dead end.
I’d probably make that the core positioning: not “AI runs your marketing,” but “your approvals teach the system how your brand should market itself.” That feels more defensible because the memory layer, scheduling intelligence, revision reasons, and campaign history all compound over time.
One thing I’d watch is the Crevaxo name. It sounds startup-ish, but if this becomes a broader marketing workflow layer for founders and lean teams, Xevoa .com would probably carry the platform direction cleaner. It feels less like a single SaaS tool and more like infrastructure for repeatable marketing execution.