Your AI orchestrator
Not a chatbot. A team.
With people, what decides the outcome is the context: what's expected, what's non-negotiable, what happened before. With AI it works exactly the same way. Growing talent and growing context are the same craft, and I've spent twenty years on the first — so I built this for the second.
Eighteen specialists, each with a written rule they don't negotiate. They don't decide, don't sign and don't publish anything on their own — but they find what nobody went looking for.
The idea and the judgement are Manu's. The eighteen are the scaffolding that holds the work while it gets polished; the scaffolding comes down at the end, and nothing goes out without going through him.
The team, in today's numbers
It doesn't run itself: someone re-measures it by hand now and then
Measured on August 30, 2026. It's a snapshot, not a live counter: someone comes back to check these numbers now and then and updates them by hand.
What's coming next
Testing ground
Two fronts still being tested internally. Nothing to show yet — once they're actually working, they'll get their own door.
Why this sits on a talent website
Growing talent and growing context are the same craft
With people, what decides the outcome is almost never the tool: it's the context. What's expected of them, what's non-negotiable, what happened before and why. That's Cultivar Talento, and I've spent twenty years on it.
With artificial intelligence it works exactly the same way. The model isn't the advantage. The advantage is whoever owns the context, works it and looks after the data — and that looks far more like caring for a team than like operating a tool.
It's one of my obsessions, so I built my own to grow it. What follows is that team, working.
And there's no platform behind it. There's a folder on Manu's computer, a file explaining how the work gets done, and one card per specialist saying what they do and what they're not allowed to do. That's the whole invention.
The rulebook
One file the whole team reads before doing anything.
It isn't a design manual: it's a list of scars. Every rule is in there because it was once done the other way and went badly. “Before deleting, look without filters” was written the day an inventory called a folder empty when it wasn't.
One rule each
Non-negotiable, with a story behind it.
Every specialist has one sentence they don't get to skip. None of them is a good intention: they all came out of a specific mistake that cost time. They're on the team page, each with the demonstration where you see it applied.
Running since March 2026. What gets decided is written down, and so is what each one learns — which is why the team makes new mistakes instead of repeating old ones.
Build your own
The starting point for building your own orchestrator, with the minimum pieces and no hand-waving. It's what used to live at this address.
Open the Starter → Who they areThe team
All eighteen, grouped by what they do, each with the rule they don't negotiate and the demonstration where you see it applied.
See the team → How it worksThe demonstrations
Seventeen situations with companies that don't exist and figures that belong to nobody. One per specialist, and three long ones where you follow the reasoning step by step.
See the demonstrations →A messy dump, and a control tower on top
This is how a real request actually arrives: a two-line message and a file with half-filled columns, unattributed rows and two different ways of counting the same thing. This is what Pinky gets — and the first thing that comes back.
Manu
Pinky, here is the H1 close for the three plants. Build me a control tower for Thursday’s board: revenue, margin, turnover and whatever else you spot. The file is what it is.
| centro | periodo | fact_eur | coste | resp | linea |
|---|---|---|---|---|---|
| VLD-01 | 2026-H1 | 2,410,500 | 1,988,100 | R. Nieto | Pavimento |
| ALM-02 | 2026-H1 | 1,905,200 | 1,702,940 | (empty) | (empty) |
| PTO-03 | 2026-H1 | 2,088,700 | 1,744,300 | M. Sanz | Revestimiento |
| ALM-02 | 2026-H1 | 0 | 0 | Ñigo Peña | Pavimento |
| VLD-01 | 2026-H1 | 0 | 0 | Iñigo Pena | Pavimento |
File preview — 5 rows out of 272.
Pinky orchestrator
Before building anything: 55 of the 272 rows have no plant and no owner, and one person shows up twice, spelled two different ways. If I put a dashboard on top of this, it looks good and it lies to your board. Here is the order:
Rex cleans the dump first: normalises the plant codes and catches
the duplicate — Íñigo Peña and Iñigo Pena are the same person.
Nash builds the tower on top, with clean data, and flags whatever
could not be attributed.- And there is one I do not decide: headcount at Almazán says 24 and the activity listing says 22. Both are correct and they measure different things. I will show you both separately and you pick which one goes to the board.
Give me a moment and I will build it.
Cerámicas Valdeluz, S.L.
Three sites · 78 people · invented companyRevenue by site
H1 2026 · solid bar = attributed · grey = no site assigned
Monthly trend
Revenue in €M · the dotted line is the same period last year
Attrition by site and tenure
Voluntary exits over average headcount · the darker, the higher
What doesn't add up
Three warnings, none of them solved by the dashboard
What I would do with this
Three recommendations, ordered by damage avoided. None of them runs on its own: each carries its confidence level and, above all, which piece of data would kill it.
The two systems disagree by two people and neither is declared the single source. Hiring on a disputed base doesn't carry an error forward: it multiplies it, because every new hire is sized against the wrong figure.
2-3 cost decisions avoided this quarter
A 40-minute meeting and one written definition
That a third record reconciling both exists and nobody is looking at it
Margin by site is being calculated on 79.8% of the dump and presented as if it were the whole. It isn't wrong: it says less than it appears to say, and in a board meeting nobody can tell the difference.
Margin by site can move by up to ±1.4 pts
Depends on why they're missing: not yet diagnosed
That all 55 lines belong to the same already-identified site — then the bias is known and bounded
Today it's four duplicates across 272 lines: 1.5%, barely noticeable. In an annual close the same mechanism builds up over twelve months and stops being noise to become a figure someone defends in a room.
4 duplicates today · the error grows every period
Low if done before the close. High if done after
Nothing: it's verified row by row. It's the only one of the three that doesn't rest on an assumption
What this dashboard does NOT claim
As important as everything above
Whoever owns the context is the one who accelerates
The tool isn’t the advantage. Whoever owns the context, works it, and looks after the data and the content: that’s who really accelerates all of these systems. If you want to go deeper into how to cultivate your talent, the whole method is one page away.
See the Cultivar Talento method →18 specialists · 1 orchestrator · 1 challenger who isn't on the team · 17 demonstrations, all built on invented data. Behind the scenes there are real engagements documented, and none of them are published here.