What We Believe
Eight principles that govern how we work — each published with what it costs us to hold.
Most consulting philosophies are safe to publish because they cost nothing to hold. "We put clients first." "We deliver excellence." Nobody has ever lost a deal over those. A philosophy is only worth publishing if it can cost you money. Here's ours, including the parts that do.
We run what we sell
We operate a real business on the systems we recommend. Todo Culebra — bookings, payments, customer conversations, dispatch, pricing, content — runs largely on AI, and our founder owns and operates it.
This is not a credential. It's a constraint. When we tell a client an approach works, we have usually already paid for the version that didn't. Our frameworks come out of an operating P&L, not a whiteboard, and the failures in our case library are our own before they're anyone else's.
We can't recommend things we haven't run. That rules out a lot of fashionable work.
No meter, no machine
If we can't measure what a system is supposed to change, we don't build it yet.
Not "we'll add analytics in phase two." Before the work starts, we establish the baseline and confirm the instrument exists and is trustworthy. A project without a meter cannot be evaluated, defended, or improved — it can only be believed in.
We turn down build work when the measurement isn't there, and propose the smaller measurement engagement instead. That is a smaller invoice on purpose.
Silence is not success
The absence of an alarm is not evidence of health. A channel with broken reporting and a channel with no customers produce identical dashboards.
So we alert on the absence of expected events, not only on the presence of bad ones. Every gauge we build must be able to answer when was I last genuinely updated — and complain on its own behalf when the answer gets old. A fallback that fails silently is indistinguishable from no fallback.
This work is invisible when it succeeds. It never demos well.
The gauge lies until proven otherwise
We assume instruments are wrong until someone has reconstructed them from primary sources. We have found our own analytics measuring tab-closing behavior and calling it checkout abandonment. We have watched a deployment tool exit green with less than half the work done.
We never let a system grade its own homework — including ours. When a platform reports on its own value, that report is a marketing artifact until it's independently verified.
We routinely tell clients their existing numbers — sometimes numbers already presented to a board — can't be relied on. That conversation is not enjoyable for anyone.
Rent the brain, own the memory
Foundation models are rented genius. Your competitor can rent the identical model tomorrow, and any advantage that lives inside the model competes to zero.
What can't be rented is the proprietary loop: your data, your feedback, your accumulated operating memory. We build so the model stays swappable and the memory stays yours. Any architecture that makes a client's advantage inseparable from one vendor's model is a liability we're helping them sign.
Vendor-agnostic builds are more work than picking one stack and marrying it.
AI theater gets killed on sight
Some AI projects exist to be seen. They demo well, survive budget season, and change nothing measurable. We name them and recommend stopping them — including when we would have been paid to build them, and including when they are ours.
The most valuable output of a diagnostic is frequently a shorter roadmap.
Exactly what it sounds like.
Adoption is a job, not a hope
A system nobody uses returns nothing, no matter how good it is. Adoption is not what happens after the project — it is a staffed, scheduled, owned part of the project, or the project has not shipped.
We would rather deliver less capability that people actually use than more that sits idle behind a login.
It makes our proposals look less ambitious than the ones they're competing against.
We'd rather lose the number than fake it
Some value genuinely cannot be measured in-period — faster iteration, capability that pays off in two years, optionality. We say plainly that those are bets, size them like bets, and refuse to dress them up as ROI.
Mislabeling a bet as a measured return is the fastest way to lose a client's trust permanently — usually about three quarters later, in a room we're not in.
An honest "this is a bet" closes worse than a confident fake number. Every time.
What we won't do
Quote a return we can't reconstruct
If we can't show the arithmetic from primary sources, it doesn't go on a slide.
Invent headcount
We are a founder-led boutique and we say so. The rigor comes from the framework, not from implied size.
Build what we can't measure
Or measure what we didn't instrument first.
Make leaving us harder
Clients should be able to walk away with the system, the source, and the memory.
Confuse uptime with accuracy
A confidently wrong AI answer is invisible to every monitor that reports the system "working."
Four questions, one discipline
Every framework we publish is one face of the same thing.
Moat & Meter
Where value actually is — and whether the advantage survives rivals renting the same models.
Read the framework →The Token Ledger
Unit economics per feature, never per invoice. A losing feature hides inside a winning one forever.
Read the framework →The Island Factory
How it actually gets built, shipped, and adopted — with adoption staffed, not hoped for.
Read the framework →The gauge discipline
Whether it worked — refusing to accept the dashboard's word for it.
Related: The Machine Audit →Companies are not short of AI enthusiasm. They are short of instruments they can defend. That gap is the practice.
If you have to prove it worked
We run AI-strategy diagnostics for operators who are being asked to defend a number. If the principles above sound like the way you'd want the work done — or if you disagree with one and want to argue about it — that's a good first conversation.