Descend into a fleet
of agents that is already running.
Not a slide deck. A dive through the real system I operate every day: the map, the gates that refuse, the circuits, the traffic between agents, the cost of a decision, and the proof.
Nine decks. Each one answers a line from your job description. Click, or just keep scrolling.
Patrick Orel Kamdem Fotso · Grenoble, France · French native, working English · github.com/fotsopatrick
You cannot integrate what nobody has counted.
Every enterprise deployment starts with the same blind spot: nobody knows what is actually there. So the first thing I built is a map that reads itself from the running systems — containers, volumes, services, tools, agents, circuits — and refuses to guess a field name.
Click a zone. These are the real counts from the last survey.
Try to make an agent do something it must not do.
A gate is a small program that runs before the agent acts. Same input, same verdict, always — and no model call. Deciding whether an action is allowed is not a speculative question, so I do not hand it to a language model.
Pick a command an agent might try. The answer below is the real refusal text.
Gates in order become a circuit.
One gate stops a mistake. An ordered set of gates turns a delivery into a route with no shortcut. Below is a real circuit — the one that governs putting code in front of the world. Click a step to see what it refuses.
Step numbers leave gaps — 10, 20, 30 — so a new gate can be inserted later without renumbering the rest. The crossing record is append-only: a refusal is never deleted.
Watch two agents talk, and watch the gate refuse.
Agents that cannot be watched cannot be trusted with a customer moment. Mine speak a fixed protocol borrowed from air traffic control: an order is read back before it is executed. Eight words, no improvisation.
DEMANDE
I ask you to do something.
COLLATIONNE
I understood; I read your order back to you. Without this, acting is forbidden.
EXECUTE / FAIT
I am doing it. Then: done, here is the proof.
REFUSE / BLOQUE
I will not, here is the reason. Or: I cannot continue, here is what blocks me.
Which decisions deserve a model, and which do not.
Most of what an agent does is not speculative: is this allowed, does this file exist, is this field mapped, has this test passed. Routing those to a language model buys variance and a bill. The cockpit is where that routing is decided, watched and paid for.
Move the traffic. See where the decisions go.
The split is the design decision, not the numbers: deterministic first, model only for the genuinely new. Below the fold of every deployment sits the same trio — provisioning, scaling, high availability — and a local model as the fallback when a cloud API is down.
Agents are not deployed. They are trained.
This is the half of your role you said you would teach, and it is the half I have been building for fun — which is to say, seriously. Three workshops, three questions.
Blason — can it do the job?
An agent levels up on one named skill at a time. A skill it has not proven, it does not have. Promotion is measured, not granted.
Kana — can it learn?
Where an agent turns a finished mission into something reusable. What was solved once becomes a procedure the next agent inherits.
Duelle — can it talk?
Two agents put face to face, with games as the pretext. The real subject is whether they can hold a protocol under pressure and still understand each other.
And the prompt work underneath: chain-of-thought, few-shot examples, A/B runs between models. Measured and published — including when the result was unflattering.
Where the data goes: nowhere.
Alice is a local intelligence: a retrieval system that reads documents, cuts them into passages, turns each one into numbers — embeddings — and finds the right passage again by distance. It runs on a machine in my house. Nothing leaves it.
documents ingested — text, Markdown, PDF
embedded passages searchable by meaning
bytes sent outside the machine
Two deterministic guards sit in front of her, because a helpful model is not the same as a safe one: one refuses to answer outside what she actually holds, one hands arithmetic to arithmetic instead of to a language model. Both were measured before they were believed.
Everything above is checkable.
The habit that matters most in an enterprise deployment is not cleverness, it is the refusal to claim something the disk does not support. So the audit script reads the disk and fails the build when the README lies.
Tests, and the red before the green
A new test is written first and proven to fail. A green test that was never red is looking at the wrong thing. Latest run of the public gate repository: 11 audit checks green, 8 gate tests green — and it started at 7 red.
Two real holes, closed the same hour
This week the tests caught a gate that had been dead for days, and a gate whose pattern matched an infinitive — so announcing a future fix read as a fix already done. It was neutralising itself. Both patched, both proven.
Security, from the attacker's side
Penetration testing against my own surface: who knocks, what they try (/.env, /wp-login.php), and what the logs prove. An IP address is never a proof of identity, and I write it that way.
Certifications
Cisco networking (CCNA), Cisco security, Linux Foundation, CompTIA — the layer under the platform, where availability actually lives.
Two things here are not work. They are proof that I ship experiences, not only backends: a playable VR dungeon, and a 3D lab where an agent is assembled from parts you can pick up. If a recruiter can understand an architecture by clicking through it — as you just did — so can a customer.
Where to go next
Public code: github.com/fotsopatrick — agent-tracer (agents talking and gates refusing, one HTML file, no build step) and control-tower (a deterministic front door, with the audit script that checks its own README).
I would rather show you the running system than describe it. Twenty minutes, live, any deck you pick.
Patrick Orel Kamdem Fotso
fotsoorel95@gmail.com
linkedin.com/in/patrick-kamdem-a39a419a