AI product
operator
I build AI products that make it all the way to a customer. Photo in, illustrated book out, through human QA, customer approval, and print fulfillment. I also built the publishing business behind 285 books and the SaaS that markets them.
The one I'd start with
StorytimeHero
A parent uploads a photo of their child. The system generates an illustrated hero, shows a personalized cover before any money changes hands, then generates the full book, holds it for human QA, and waits for the customer to approve it before a single page goes to print.
Next.js, Payload, Postgres, Stripe, Gelato, and a separate generation worker. Most of the work sits after the model returns an image, in deciding whether it is good enough to send to a child.
Also live, also mine
All three are running in production right now. Open any of them.
PubPod Live SaaS
Multi-tenant marketing platform for independent publishers. Public pricing, self-serve onboarding, publisher terms, privacy, and compliance surfaces. Books, readers, campaigns, consent and suppression, reader portals. Runs on AWS, delivering through Amazon SES with bounce, complaint, and suppression handling.
pubpod.co →
StripeFit Live
1,679 published running-gear guides from an automated content system, with disclosure on every page, a published methodology, and a hard editorial boundary keeping gear advice out of medical advice.
stripefit.com →
Med Spa Momentum Early stage
AI services for aesthetic practices on GoHighLevel: lead capture, follow-up, no-show recovery, dormant-patient reactivation, referrals. Includes a 173-page published book. New, and no traction is claimed.
medspamomentum.ai →The rest of the evidence
Seven imprints, 285 books. Each one takes a click to check.
Writing
Things I worked out by building, written up so the next person does not have to.
By the numbers
on Amazon
and operated
with disclosure
in production
finance
How I work
Most of the work happens after the model returns something.
State before features
Take a workflow nobody has modeled and make its state explicit first. What exists, what is pending, whose turn it is, and what happens if it fails.
Humans at the consequential edge
Automate the middle, never the moment where a mistake reaches a person. Decide deliberately where a human has to stand, then build the system so they cannot be skipped.
Fail closed
When the evidence is incomplete, block the release. A model's confidence is not a test result.
Say where the edges are
Draw the line the automation must not cross, in writing, then enforce it in the system rather than trusting good intentions at volume.
Teaching is how I check my own understanding
If I cannot hand a thing to someone who has to act on it, I do not understand it yet. That test has shaped how I build, document, and hand off work for seventeen years.
Find the representation that carries the signal
Prose can specify content but not spatial arrangement. For worksheet generation I hand the model a standalone ASCII layout as the controlling composition reference, because the arrangement is where the pedagogy lives.
Work
- Promoted to Financial Advisor in 2019; built and ran the San Francisco practice for six years.
- Sourced and placed private-market investments with accredited clients: pre-IPO SPVs, private placements, and private deals, covering diligence, suitability, and investor education on illiquid, high-risk positions.
- Built LinkedIn client-acquisition systems and end-to-end webinar funnels during the pandemic.
- Series 7, 63, and 65 (earned, not active).
Also
Five full-distance Ironman races, eight 70.3s, four Olympic-distance races, and a Kona qualification.
Ten years in restaurants before any of this, from casual dining at sixteen to three summers of fine dining at Chatham Bars Inn and front of house at the Sonnenalp in Vail. Reading what someone actually needs and delivering it under time pressure is a job I have had in one form or another ever since.
Lived in Spain and Argentina, traveled in more than 35 countries. Two years of college coursework, no degree.