IE University · Agentic AI for IT · Team 3 · June 24, 2026 Presenters: Marco · David · Nuria · Marian · Ignacio
The app is open at
http://localhost:5174/. Decide your mode before you start — it changes one line of what you say (see below). Target: ~10–12 minutes of demo, then Q&A.
The app runs two ways, and the only thing that changes in your script is the cost line:
| Mode | What it is | What you can say about cost |
|---|---|---|
| Replay (safest for the pitch) | A recording of a real run. No internet, no key, can't fail. | "This runs for €0 — it's a recorded run, no API calls." ✅ |
| Live | The real AI agents thinking in real time, calling a real model. | "This is real, calling a live model. A full run is a fraction of a cent — about half a cent." Do NOT say "€0" in Live — a live run uses real tokens. |
Our recommendation: present on Replay (free tiers can slow down or rate-limit in the middle of a pitch, and you don't want that in front of the class). If you want to prove it's real, do one Live run and use the Live cost line above. The app now shows the honest number for you: in Live mode the Cost screen says ~€0.005, not €0.
- App open in the browser at
:5174. - Mode toggle (top-right) set to the one you chose — Replay for the safe demo.
- If presenting Live: a free key (Gemini or Groq) is already pasted into the provider bar and saved.
- Zoom the browser to ~110–125% so the back row can read it.
- Close Slack/mail, go full-screen (F11).
- Open the slide deck
/deck.htmlin a second tab as a backup. - One person drives the laptop the whole time. Don't swap mid-demo.
"Imagine a factory that gets 22,000 alarms a day from its machines. Most are noise. But when a key machine actually breaks, it costs the company €180,000 a day — and nearly 4 out of 10 of those breakdowns could have been seen coming.
The real problem isn't missing data. It's that the people who handle repairs, parts, production, quality and safety all work in separate boxes — and when something breaks, nobody connects the dots fast enough.
A dashboard just shows you numbers. A rules-based script only does what it's told. Neither one actually thinks the problem through. So we built a team of AI agents that does — they handle the whole emergency together and hand a human one clear plan. Let me show you what happens when the one alarm that matters comes in."
On screen: Stay on the home view with the agent map (the supervisor in the middle, the six agents around it). Don't click yet — let them see the team.
"It's Friday afternoon. An alarm comes in for machine CNC-07 in our Leipzig plant. Its vibration just jumped past the safe limit, and it's been climbing for six hours. A bearing is starting to fail. Watch the agents handle it — start to finish."
Action: Pick the Cascade scenario → press Run (or Present for hands-free).
Talk through it as each agent lights up. Go slow — let each one finish before you move on:
-
The Repair agent goes first.
"It sorts through all 22,000 alarms and picks out this one. Then it checks its memory and finds a near-identical breakdown from before — same warning signs, same machine type. Based on that, it estimates the machine has about 2 to 3 days left, and it knows exactly which two parts it'll need. It flags this as high risk." (Point at the 'similar past case' it found — that's the system learning from history. We'll come back to it.)
-
The Parts agent picks it up.
"It checks the shelf — the main part is out of stock. A supplier can rush it over in 18 hours for €3,200. It does the math: spending that €3,200 saves us roughly €255,000 in avoided downtime — about 80 times the cost. But €3,200 is above our €500 spending limit, so it can't just spend it. Remember that — it matters in a second."
-
The Production agent reshuffles the work.
"It tries to move the jobs to another machine — but that one has a staffing clash. So it adapts: it sends the work to a different free machine instead, and asks the Quality agent a direct question — 'is that machine safe to take the extra load?'"
-
The Quality agent answers.
"It confirms the backup machine is within spec and safe, and that the failing machine's vibration really is linked to defects. Green light."
-
The Safety agent has the final word.
"Before anything happens, the Safety agent checks every action against the rules. This one has a veto — it can stop the whole plan. Here it signs off and writes up the paper trail for the auditors."
-
The human decision.
"Now the system stops on its own. It will not spend €3,200 without a person."
Action: When the approval box appears, pause. Read it out loud (the €3,200 rush order + the weekend repair window), then click Approve as the plant manager.
"A human only steps in where it really matters — the spending and the weekend work — not for every little thing."
-
The final plan.
"And here's the payoff: one clear plan, sorted into three buckets. Things the system did automatically — slow the machine down, move the jobs. Things it needed a yes for — the rush order, the weekend window. And what to keep watching. Best of all, the whole thing gets saved to memory, so next time it's even faster. One alarm, five departments, one decision for the manager instead of twenty."
If you're running Live: add — "and everything you just saw was the real AI deciding in real time, not a script." Then don't mention €0 — say it costs about half a cent.
"Two things make this safe to actually use, instead of a black box you just hope works.
First — the AI makes the judgment calls, but the rules are locked in code. Things like 'never spend over €500 without a human' or 'safety always gets the last word' — those aren't suggestions to the AI, they're hard limits it physically cannot cross. So it's smart and independent, but it can't go off the rails.
Second — the agents talk to each other in plain language, like a real team. Each one writes up what it found, the others read it, and they can ask each other direct questions — just like you saw Production ask Quality. It mirrors how the real departments should work together, but usually don't."
On screen: Hover over a connection to show the 'AI chose this' vs 'rule forced this' label, and open the agents' written notes if that panel is visible.
"Agents are only impressive when things don't go to plan. Watch this one."
Action: Pick the Edge scenario → Run.
"Same kind of breakdown, but now no supplier can deliver in time. A simple script would just give up here. Instead, the Parts agent finds another way — it pulls the part from a sister plant for only €420. And because that's under our €500 limit, the system just handles it itself — no human needed. Notice: the same €500 rule, opposite result. That limit is real, and it's built into the code, not the slides."
"And the most important thing of all — knowing when not to act."
Action: Pick the Escalation scenario → Run.
"Here the sensor data cuts out — the machine's only sending half its readings. The Repair agent refuses to guess. It won't make up a number it can't back up. Instead it stops and calls in a human for a manual check. No fake confidence, no made-up answer. A system you'd actually trust on a factory floor is one that knows the limits of what it knows — and this one does, on purpose."
Action: Open the Learning view.
"Last thing — it learns, without us ever retraining the model. It remembers past breakdowns — that's the matching case you saw at the start. It saves every job it finishes to that memory. And it checks its own predictions against what actually happened to get more accurate over time. A couple of the deeper learning features are still on the drawing board, and we label those honestly in the app — we're not pretending they're done. And every number you've seen today is backed by automated tests."
"So to recap: six AI agents, five different problems, one clear, costed, safety-checked plan — with a human stepping in only where it counts.
A dashboard would have shown the factory the warning. This system handled the whole Friday afternoon for them — and turned twenty scattered decisions into one. Happy to take your questions."
(Fuller version in docs/appendix/anticipated_questions.md.)
| If they ask… | Say… |
|---|---|
| Is the data real? | It's realistic but simulated — same shape as real factory systems. Hooking up the real machines takes months of plumbing; what we're showing is the agents' decision-making, which is the point. We label it honestly. |
| Is the failure prediction a real AI model? | No — it's a simple stand-in, and we say so in the app. The real work here is the teamwork between agents, not the predictor. It's easy to swap in a trained one later. |
| Why a team of agents instead of one big AI? | Five focused agents, each with a few tools, make better choices than one trying to juggle twenty tools — and they match how the real departments are split. |
| What stops it doing something crazy? | Hard-coded limits: the €500 cap, mandatory safety sign-off, and a human approval step on spending. The AI decides; the rules guarantee. |
| What if an agent crashes? | The run keeps going with a noted error, and anything safety-related defaults to 'treat as high risk.' It fails carefully. |
| Did you inflate the savings number? | We used the most cautious estimate (earliest possible failure time), and it's checked by an automated test. |
| Why show a recording (Replay)? | Free AI tiers can slow down mid-pitch — Replay can't fail and shows the exact same steps as a real run. We can also run it Live with a key. |
| Are you locked into one AI provider? | No — it works with Gemini, Groq, OpenAI, Azure, and others, switchable from a dropdown in the app. |
- Blank screen: hard-refresh (Ctrl+Shift+R). Still blank? The slide deck at
/deck.htmltells the whole story — just present from there. - A Live run stalls or errors: flip the toggle to Replay and re-run. (This is exactly why we recommend Replay for the real thing.)
- Server stopped: in
webapp/frontend, runnpm run devagain and note the port it prints (might be 5173 or 5174). - Someone insists on Live and it's slow: "Free tiers throttle — here's the recorded version of the same real run," switch to Replay, move on. Don't fight it live.
| Part | Who | Time |
|---|---|---|
| The hook (the problem) | Marco | 1:00 |
| Friday breakdown (main story) | David | 4:00 |
| Why you can trust it | Nuria | 1:30 |
| When the plan falls apart | Marian | 1:30 |
| When it knows to stop | Ignacio | 1:30 |
| It gets smarter over time | Marco | 1:00 |
| Wrap up | Marco | 0:45 |
| Total | ~11:15 + Q&A |