25. First Principles

In a tense boardroom showdown, Bells calls Jude’s bluff on his wildly ambitious Yamamoto proposal. But when Yen walks in, the battle shifts from business to something far more personal.

Bells sent the invite early morning.

Topic: Yamamoto Business Case Review

To: Jude Larssen

Time: 10:30

Where: Conference Room 3B

He’s accepted immediately. Had been pestering her to sign off for days. Once he’d clicked “Yes,” she added four more people with a note:

· Mark, Head of Finance

· Jack, Lead of Engineering

· Keith, Head of Legal

· Jason, all-things-HR

“As per our earlier discussions”

Then, a separate 11:00 invite to one extra person. Just to keep him on edge.

Bells and the four men arrived at 10:25. She sat at the head, made a minute of small talk, but the air was tight.

At 10:37, he walked in. Jude Larssen. Like he owned the place.

Which technically, he did.

His state always betrayed the day’s battle load. Today: brutal. Hair rumpled, shirt crumpled, eyes sunken. Confidence intact.

He scanned the room.

Jude: - This what I’ve been summoned for?

Bells: - Morning Jude.

The men nodded in courtly acknowledgment. He took the opposite end from her, waiting.

Bells: - Your Yamamoto business case reads like speculative fiction. And not a particularly good one at that.

He leaned back in his chair.

Jude: - So this is an intervention.

Bells: - First, how exactly are you planning to cover 50% of their European sites in year, when Mitsubishi achieved 25% coverage and that was with IBM and Siemens holding their hands, providing the software and IoT layers.

Mark (finance): - Secondly, your payment terms are front-loaded. Twenty percent of Year 1 value in Q1? That’s a £10 million cashflow gap.

Jude: - Innovation’s not cheap, Mark.

God. Always so arrogant. Bells thought.

Jack (engineering): - Third, your delivery timescales, even for just the software layer – triple that and you’re still running lean.

Keith (legal): - Legislatively, predictive models will produce false positives. If Yamamoto halts operations for a false alarm, we could face serious legal exposure.

Bells: - Millions in penalties. You predict something breaks, it turns our erroneous, but they stop operations to fix.

Jason from HR pretended to follow and sipped his tea. Jude clasped his heads together in front of his face.

Jude: - You’re done?

Bells: - Do you want to explain exactly how any of this is supposed to work?

Jude: - Alright, let’s talk. Since you all are feeling so… conflicted.

Took a pause and then a sip from his water bottle.

Jude: - Mark. Yes, I pulled forward the revenue. And yes, they usually pay late. But not on this. They’ve got ¥40B in receivables this quarter, a looming capex freeze, and political heat on their board. They need it closed early. It helps their optics. Helps them restructure under the radar.

Bells’ gut tightened, this reeked of inside knowledge. Yen again.

Mark: - Conjectural.

Jude: - It’s not, but to offload our risk profile, I’ve mortgaged some personal assets to part underwrite if need be.

The room has gone quiet.

Bells: - You have that much lying around?

Jude: - My apartment, you’ve seen it.

Her pen nearly snapped in her hand.

And then Yen walked in. Ten minutes early. Smiled straight at Jude. Of course she did. Bells wondered if she’d still be smiling in a minute.

Jude: - Yen. Nice of you to join us.

Big boyish smile. And then, he stood up. Approached the whiteboard, picked up a pen.

Oh no. Not the whiteboard.

Jude: - As to your other queries. Let’s do it from first principles.

He started writing.

Assumptions

1. Insufficient capacity vs. bigger players

2. Short timescales

3. Risk of disruptions from false positives

Jude: - Now, one.

He sketched roughly an outline of Europe. Then circling each country as he spoke…

Jude: - Germany, Belgium, France, Netherlands and Sweden plants. 24 altogether, slightly over the 50% target actually, Bells. All running preventative maintenance framework already. Meaning – the IoT sensors are in place. Of course you don’t read it in the news, they love a disaster story.

Bells: - So what are you saying? We only ship the software layer?

Jude: - Precisely. Sensors are already talking to the cloud. We put the algorithms to work to enable shift to predictive. We’re not dragging the laggards along, but pushing the leaders ahead. Mitsubishi were neck-deep in reactive. Yamamoto’s half-way there.

Bells tilted her head, considering.

Jude crossed one out.

1. Insufficient capacity vs. bigger players

Jude: - Now. Two. Jack. What is our competitive advantage?

Jack (engineering): - Top engineering talent?

Jude: - Funny.

Yen: - Leadership?

Jude turned to look at her, grinning. Bells took a deep breath. Poor naïve Yen, she’d fallen for him.

Jude turned back and wrote on the board again.

GB2523978

GB3432069A

GB3713454

Jude: - GB2523978 patent – comms protocol for Argus, GB3432069A.

Jason (HR): - What’s that?

Jack (engineering): - Argus is our platform for managing assets that connect with third-party sensors through Azure Cloud…

Jude: - … for real time data collection. And then GB3713454. Pay attention Keith. Our hybrid model utilising LSTM networks and Bayesian inference, incorporating machine maintenance records to determine priors.

He started drawing failure distribution curve, captioned it as “prior”.

Jude: - Likely Weibull.

Then an arrow into a sketch of a cluster diagram. Captioned “LSTM”.

Jude: - Roughly.

Then an arrow pointing to a number: 97%.

Jude: - 97% model precision. False positive rate lower than competition.

Jack (engineering): - Which sounds dangerously like overfitting.

Jude: - 95% is industry standard Jack. And this has been proven in the Belgium pilot. Inference datasets performance.

Bells: - What’s the uncertainty on that? With small n it can swing.

Jude: - Fair. In the Belgium pilot we labelled 1,300 events and ran live inference. 170 true positives. Five false positives. Precision on the failure class was 97%

Started writing on the board again.

\(\hat{p} \pm z_{\alpha/2} \sqrt{\frac{\hat{p}(1 - \hat{p})}{n}}\)

Jude: - Confidence interval on precision, that’d be the estimate plus/minus standard error times z-score at alpha halved. Assuming normal, n=175, and alpha 5%.

\(0.97 \pm 1.96 \sqrt{\frac{0.97 \times 0.03}{175}} \approx [0.945, 0.995]\)

Jack (engineering): - So market mean at lower bound.

Jude: - Correct.

She hated when he quoted pilot data. Was difficult to counter.

Keith (legal): - You are still left with the mean 3% risk.

Jude: - We build liability caps into contract structure and I’m in the process of negotiating insurance coverage based on our risk profile.

Keith (legal): - And you didn’t think to run this by me?

Jude: - I need this within days, not months.

Bells mouthed “sorry” to Keith. Not the first time she had to act as a buffer between Jude and his staff.

Jude crossed out assumptions (2) and (3) of the board.

2. Short timescales

3. Risk of disruptions from false positives

Jack (engineering): - Wait, not sure using Argus exactly speeds up delivery sufficiently.

Jude: - Argus’ plug and play. Cloud native, API-first. The Belgium pilot was our proof of concept.

Jack (engineering): – Pilot’s one thing. Scaling to 24 sites, with varied sensor hardware and network quality? That’s extra dev time and unforeseen bugs.

Bells: - Isn’t that one of the arguments you put to me before we met Yamamoto-san for dinner, Jude? “Try implementing predictive maintenance at Osaka. Or Leeds. See how far “smooth” gets you there” you said.

Jude scratched his head.

Jude: - Okay. But for Y1, the 50%, we are cherry picking sites. We’re not going for Leeds. Certainly not Osaka. We identify sites most similar to Belgium.

Bells had a small smile on her face. Like the argument was not entirely persuasive.

Jack (engineering): - Even then, Belgium took 12 weeks in full swing. 24 sites × 12 weeks = 288 weeks. And you’re promising 52. Do you want me to put that on the board?

Jude was annoyed now. Bells could tell. She liked Jack. He could stick it to Jude on occasion.

Finally Jude spoke.

Jude: - No. But watch this.

He started whiteboarding again. Furiously.

Jude: - Deployment… timescales… are not… linear Jack.

\(\begin{aligned} &\text{Deployment time:} \quad T(n) = T_{\min} + (T_0 - T_{\min}) e^{-\lambda n} \\ &\text{where} \quad T_{\min} = 2,\ T_0 = 12,\ \lambda = 0.2 \\[6pt] &\text{Examples:} \\ &T(5) = 2 + 10 e^{-0.2 \cdot 5} = 2 + 10 e^{-1.0} \approx 5.7 \ \text{weeks} \\ &T(10) = 2 + 10 e^{-0.2 \cdot 10} = 2 + 10 e^{-2.0} \approx 3.4 \ \text{weeks} \\ &T(24) = 2 + 10 e^{-0.2 \cdot 24} = 2 + 10 e^{-4.8} \approx 2.1 \ \text{weeks} \\[6pt] \end{aligned}\)

Jude: - We learn as we go along, at lambda 0.2. Each site takes less and less time until site 24 takes just over 2 weeks.

Mark (finance): - You could have just used the standard learning curve, Jude. T equals T-one times N to the negative-B. Basic MBA stuff.

Jude: - Learning curves assume linear log relationships, Mark. That's business school crap for manufacturing widgets. This is exponential knowledge transfer.

Bells: - And what’s that in total Jude?

Jude: - What?

Bells: - You’re exponential decay model. What does that add to timescales-wise?

He went to the board again.

Jude: - You have to… sum… the geometric series.

\(\begin{aligned} &\text{Total serial:} \quad \sum_{n=1}^{24} T(n) = 48 + 10 \sum_{n=1}^{24} e^{-0.2n} \\ &\sum_{n=1}^{24} e^{-0.2n} = e^{-0.2} \frac{1 - e^{-4.8}}{1 - e^{-0.2}} \approx 12.8 \\ &\text{Total} \approx 48 + 128 = 176 \ \text{weeks} \\[6pt]\end{aligned}\)

Jack (engineering): - Uh. I mean. That’s just over three years. Like I said - triple the timescales, you’re still running lean. Now mathematically proven.

Bells chuckled. He’d just shown himself reckless.

Mark (finance): - Ergo, this project is impossible in a year.

Jude smirked suddenly.

Jude: - Which is why we don't deploy serially. We deploy in parallel streams.

He went to the board again.

\(\begin{aligned} &\text{Parallel (3 teams):} \\ &\text{Team A:} \quad \sum_{n=1}^{8} T(3n-2) \approx 72 \ \text{weeks} \\ &\text{Team B:} \quad \sum_{n=1}^{8} T(3n-1) \approx 64 \ \text{weeks} \\ &\text{Team C:} \quad \sum_{n=1}^{8} T(3n) \approx 58 \ \text{weeks} \\[6pt] \end{aligned}\)

Jude: - Three deployment teams. Each follows the exponential learning curve independently, but they share knowledge across streams. Team A: sites 1,4,7,10... Team B: sites 2,5,8,11... Team C: sites 3,6,9,12...

\(\text{Cross-team learning:} \quad 72 \times 0.75 = 54 \ \text{weeks}\)

Jude: - All teams finish within 72 weeks. But with knowledge transfer between teams, we compress this further. Cross-team learning coefficient reduces delivery time by additional 25%.

Everyone was quiet for a moment. Then Jack.

Jack: - This could… Actually work.

Jason (HR): - I still don’t get it.

Jude looked at Jason.

Jude: - Remind me, why are you here?

Jason went speechless.

Bells: - Staffing matters. Leave commitments. Burnout risk?

Jude: - No one takes leave during key sprints, especially Jack’s team - slow as they are.

Jack (engineering): - We’re not ‘slow’, we are overtasked and underpaid.

Jude exhaled.

Mark: - Okay, I see where you are coming from. You’re still just over a year but maybe not sci-fi after all.

Then suddenly.

Yen: - You’re brilliant Jude.

She said with such eager, honest admiration. It rattled Bells. No one did that, not to his face. They had to keep that ego in check. Especially Bells. She looked at Jude. Expected him to stand taller but his shoulders loosened. A soft smile on his face. Really? That’s all it took? Some flattery?

Jude: - Any more questions or can I get your sign off now Bells?

Bells: - Well, you need to put some of those figures in the document.

Jude: - Thought that’d be patronising.

Bells shook her head. The arrogance was back.

Bells: - Fine. I will do it. And perhaps adjust your timescales by a few weeks. And we might need to hire some engineers.

Jack (engineering) : - Oh yes. That would help.

Jude: - Whatever. Just don’t blow up the cost schedule.

Chairs scraped, papers gathered. They filed out.

Only Jude and Yen lingered, her leaning in, laughing at something he said.
Bells looked through the glass outside. Just noted the angle of their heads, the distance between them.
Oh dear.

---

Research:

https://www.sciencedirect.com/science/article/pii/S2667305325000274

This is a technical one. Very long. I did think about trimming this down but if I'm honest, I enjoyed every single line so didn't :)