Product leader · AI, data, life sciences · Vienna

Efthymios Vogiatzis: product leadership in AI and life sciences.

A small discrepancy is never small.

I trained as a pharmacist, where you treat the tiny thing as a real problem until it proves harmless. That habit turned out to be worth more in a messy dataset than I expected, and it is most of what I do now.

Read how I got here
Portrait of Efthymios Vogiatzis in a laboratory

The operating thesis

Intelligence is the commodity now. The system is the edge.

What the counter taught me

From the counter to the cloud

The first job I ever had was behind a pharmacy counter. A prescription would come over, and the work was checking the dose, the interaction, the name that reads almost like another name. The interesting part was never the obvious one. It was the almost.

That habit, treating the tiny thing as a real problem until it proves harmless, became the only useful skill I carried out of the building.

You treat the tiny thing as a real problem until it proves harmless.

The military pharmacy year

After the counter came a uniform. From 2014 to 2015, as an officer in the Hellenic Army Pharmacy Service, I spent nine months running a base pharmacy, controlled substances, cold chain, an audit trail that had to hold under a surprise inspection. There is no room there for probably fine. You count, you sign, you reconcile. The stock either matches the ledger or someone has a problem. Nine months turned the pharmacist's instinct into a discipline: never assume, always verify, leave a trace.

Then came the molecules, on a screen

Years later, the same instinct pulled me into DeepLab, a health-tech startup using AI for virtual screening in drug discovery. The premise is elegant: instead of testing millions of candidate molecules in the wet lab, you let a model rank them first, and only the most promising few earn a pipette. It compresses years of bench work into days, when the data underneath is clean.

That is where the pharmacist's eye earns its keep. A training set with subtly mis-labelled assays, a target annotation that has quietly shifted in the public databases, a molecule that looks almost like a hit until you read the stereochemistry, these are the failure modes that no accuracy metric catches. The model is only as honest as the rows it was raised on.

For more than a decade since, I have been leading AI and machine-learning products from concept into production, across life sciences, financial services and industrial sectors. The geographies have changed. The discipline has not.

The work that holds up is rarely the model. It is the data the model was built on, the decision the model is feeding into, and the room where that decision gets made. Most of what I do sits in the seam between those three.

Field reading · the drift

The number on the slide is not the number in the world.

Most reports do not break. They simply stop describing the thing they were built to describe. The gap opens quietly, and the room keeps deciding from the old picture.

60708090Week 1 · Reported 82Week 1 · Actual 82 (gap 0)Week 2 · Reported 83Week 2 · Actual 82 (gap 1)Week 3 · Reported 82Week 3 · Actual 81 (gap 1)Week 4 · Reported 84Week 4 · Actual 82 (gap 2)Week 5 · Reported 83Week 5 · Actual 80 (gap 3)Week 6 · Reported 84Week 6 · Actual 81 (gap 3)Week 7 · Reported 85Week 7 · Actual 79 (gap 6)Week 8 · Reported 84Week 8 · Actual 78 (gap 6)Week 9 · Reported 85Week 9 · Actual 76 (gap 9)Week 10 · Reported 84Week 10 · Actual 74 (gap 10)Week 11 · Reported 85Week 11 · Actual 71 (gap 14)Week 12 · Reported 86Week 12 · Actual 68 (gap 18)Week 13 · Reported 85Week 13 · Actual 64 (gap 21)Reported85Actual64Week 1Week 13
Reading the gap. The dashboard kept reading 85. The thing it was meant to measure had quietly walked to 64 over thirteen weeks, an 18-point divergence invisible to anyone watching only the reported line.Source. Stylised composite from three operational KPI reviews (life sciences and industrials, 2020–2024). Values normalised to a 0–100 index; weekly cadence preserved from the source dashboards.
0×25×50×75×100×Source1×Source · 1× the cost to fix vs. catching it at the sourcePipeline4×Pipeline · 4× the cost to fix vs. catching it at the sourceWarehouse12×Warehouse · 12× the cost to fix vs. catching it at the sourceDashboard30×Dashboard · 30× the cost to fix vs. catching it at the sourceBoardroom80×Boardroom · 80× the cost to fix vs. catching it at the sourceWhere the discrepancy is found
The compounding bill. Same error, five rooms. Caught at the source it is a ticket; caught in the boardroom it is a decision already made, roughly 80× the remediation cost.Source. Cost multipliers adapted from IBM Systems Sciences Institute defect-cost research and calibrated against incident post-mortems from two enterprise data programmes I led (2021–2024).

Field reading · the compounding bill

A tiny thing is a real problem until it proves harmless.

Discrepancies do not stay the same size. They acquire followers, decisions, slides, commitments , and the cost of unwinding them scales with how far downstream they have travelled.

What I write about

Three things I keep coming back to.

  • 01

    How data quietly stops telling the truth

    Drift, contamination, silent schema changes. The work is catching it early, before a decision is already in motion.

  • 02

    Why the human side of a change decides whether it survives

    Models do not fail in production. Trust does. The interesting problem is rarely the algorithm.

  • 03

    What a pharmacist's eye notices that a strategy deck misses

    Names that read almost like other names. Doses that look almost right. The almost is where the work is.

Efthymios Vogiatzis at a city office window at dusk

Field reading · the listening curve

Models do not fail in production. Trust does.

Accuracy is cheap. A decision that actually changes is expensive.

0%25%50%75%100%Pilot · Model accuracy 92%92Pilot · Decisions actually changed 38%38PilotMonth 3 · Model accuracy 90%90Month 3 · Decisions actually changed 44%44Month 3Month 6 · Model accuracy 88%88Month 6 · Decisions actually changed 41%41Month 6Month 12 · Model accuracy 87%87Month 12 · Decisions actually changed 28%28Month 12Model accuracy (%)Decisions actually changed (%)
What the chart says. Held-out accuracy stayed within one point of 90% for a full year. The share of eligible decisions the model actually moved fell from 38% at pilot to 28% by month twelve, a 26% relative drop in influence while technical performance was steady. Source. Composite of four enterprise AI rollouts I led or advised on between 2019 and 2024 in life sciences, insurance and industrial operations. Figures anonymised, ranges normalised to a 0–100 scale. n ≈ 1,200 tracked recommendations per programme.

Why the name

VE1 Advisory

In aviation, V1 is the speed past which you no longer stop the takeoff. From here, you commit, and you fly.

Most decisions in a business feel like a cliff. They are not.

VE1 reads as vee-one. The initials are mine , Vogiatzis, Efthymios, but the sound is borrowed from a cockpit. Every commercial pilot listens for it on every takeoff. It is the moment the aircraft has enough energy that pulling the throttles back would be more dangerous than continuing into the air. So they continue. Calmly. By plan, not by feel.

The job of a good advisor is not to make the decision dramatic. It is to make the moment of commitment quiet.

That is the work I want VE1 Advisory to be known for. Not the slide that announces a transformation. The months of small calls before it, the data review that catches a discrepancy, the workshop that names the real risk, the org redesign that someone was too polite to suggest. By the time the decision is taken, the noise is gone.

Pilots don't call V1 a leap of faith. They call it a number. Computed for this aircraft, this weight, this runway, this wind. That is the shift I want to make for the teams I work with. Less courage, more calculation. Less dramatic, more altitude.

01

Compute the number

Before commitment, the data, the units, the assumptions are reconciled. The number on the slide is the number that survives a board question.

02

Name the commitment

We make the moment of no-return explicit. What is being decided, by whom, against which trade-offs, and what stops mattering after this.

03

Fly the aircraft

After V1, the work is execution, not debate. We stay with the team through the climb, the first quarter, the first review, the first surprise.

Field notes · twenty lines from the cockpit

The lines I keep coming back to.

Twenty short observations, grouped into five chapters. Each one is a working hypothesis I have stress-tested in a laboratory, a trading floor, a board room, or a factory. Together they are the closest thing I have to a method.

01

Chapter I · commodity & edge

Intelligence is the commodity now. The system is the edge.

Every team can rent the frontier model. Almost no team can rent the workflow, the data discipline, and the trust around it. That second layer is where advantage actually lives.

I keep meeting leaders who have just signed a six-figure contract for a model and are sure the hard part is over. It is the moment I get worried for them. The model is the engine. The aircraft is the data, the controls, the crew, the maintenance log, and the room where the captain makes a call. The engine is for sale. The aircraft is not.

Anyone can buy the engine. Few rebuild the aircraft around it.

  • 01.01

    Intelligence is the commodity now. The system is the edge.

  • 01.02

    The model is rented. The system is earned.

  • 01.03

    Everyone can rent the model. No one can rent the system.

  • 01.04

    The frontier is for sale. The advantage is not.

  • 01.05

    The model sets the ceiling. The system sets the altitude.

  • 01.06

    The intelligence is a utility. The judgment is the work.

  • 01.07

    Anyone can buy the engine. Few rebuild the aircraft around it.

02

Chapter II · the pharmacist's eye

A small discrepancy is never small.

The pharmacist's habit is to treat the tiny thing as a real problem until it is proven harmless. That instinct, moved into a data team, is worth more than most algorithms.

On the counter, the discrepancy is a dose that ends in 5 instead of 0.5. In a model, it is a unit silently converted at ingestion. In a board pack, it is a footnote nobody reads. The job is the same in all three rooms, notice the almost before it walks out of the door.

The work is in the almost.

  • 02.01

    A small discrepancy is never small.

  • 02.02

    The work is in the almost.

  • 02.03

    The interesting part was never the obvious one.

  • 02.04

    A tiny thing is a real problem until it proves harmless.

03

Chapter III · drift & the number

Most reports do not break. They quietly stop being true.

Failure in analytics is loud and rare. The dangerous mode is drift: the report keeps rendering, the chart keeps trending, and the world it described has walked off without it.

A team I worked with watched their flagship KPI hold at 85 for three quarters. The model was fine. The pipeline was fine. The world, when we finally measured it directly, was at 64. Nothing had failed. Everything had drifted. The cost of those three quarters was, by then, larger than the project that produced the dashboard.

The dashboard read 85. The world had walked to 64.

  • 03.01

    Most reports do not break. They quietly stop being true.

  • 03.02

    The number on the slide is not the number in the world.

  • 03.03

    The dashboard read 85. The world had walked to 64.

04

Chapter IV · trust

Models do not fail in production. Trust does.

I have never seen a deployed model fail catastrophically. I have seen rooms quietly stop using one. The shape of the failure is human, not technical.

A model that is 90% accurate but cannot show its working is, on the third Tuesday of the rollout, a model nobody opens. Accuracy is a property of the algorithm. A decision that actually changes is a property of the relationship between the algorithm and the person reading it. The second one is what we are actually paid to build.

Accuracy is cheap. A decision that changes is expensive.

  • 04.01

    Models do not fail in production. Trust does.

  • 04.02

    Accuracy is cheap. A decision that changes is expensive.

05

Chapter V · the V1 commitment

Past V1, you commit. Then you fly.

In a cockpit, V1 is not bravery. It is a number, computed for this aircraft, this weight, this wind. The good advisor's job is to do the same arithmetic for a business decision, so the moment of commitment is quiet.

Most boardroom drama is a sign the homework was not done. By the time the decision is in the room, it should already be inevitable: the data has been reconciled, the trade-offs are named, the people are aligned. The signature is the lightest part of the day. After that, the work is altitude, not debate.

The job is to make the moment of commitment quiet.

  • 05.01

    Past V1, you commit. Then you fly.

  • 05.02

    Less courage. More calculation.

  • 05.03

    The job is to make the moment of commitment quiet.

Selected work

Where the eye was sharpened.

  • Thermo Fisher Scientific

    Product Line Owner, Chromatography & Mass Spectrometry

    AI from prototype to production, on data foundations you can trust.

    2023–2026

  • Retail pharmacies

    Strategic advisor, operations & M&A

    A long-running thread: advising independent pharmacies and small groups on operations, succession and consolidation.

    since 2015

  • Renewable energy (solar)

    Developer & operator, solar park projects

    A parallel thread: building and running solar parks. Long horizons, real assets, patient capital.

    since 2012

  • Takeda

    Commercial Insights & Analytics Lead

    Orchestrating data across stakeholders to inform real decisions.

    2021–2023

  • GLG (Gerson Lehrman Group)

    Council Member, freelance expert network

    On-call expert for investors and operators on pharma operations, retail pharmacy and life-sciences data.

    2014–2021

  • DeepLab

    Healthtech advisor, AI-driven virtual screening

    Helped an Athens-based health-tech turn computational chemistry and ML into a virtual-screening engine that narrows millions of molecules down to the few worth a wet-lab look.

    2021

  • Johnson Controls

    Senior Analyst, Global Information Security

    Risk reporting where the wrong number is worse than no number.

    2020–2021

  • Swiss Re

    Digital & Smart Analytics

    Analytics that change how a business decides, not just how it reports.

    2019–2020

  • Bioiatriki Group

    Diagnostic Center Supervisor, Athens

    Running a diagnostic centre where throughput and quality cannot trade off, every sample is somebody’s morning.

    2018

  • Hellenic Army Pharmacy Service

    Officer, military pharmacist

    Nine months of mandatory service running a base pharmacy: controlled substances, cold chain, audit trail. Where “never assume, always verify” became muscle memory.

    2014–2015

  • Community pharmacy

    Pharmacist, family pharmacy

    Counter, counsel, inventory, regulation. The first place I learned that a small discrepancy is never small.

    before 2014

Podiums, pages & quiet service

Where I speak, write and lend my name.

  • 2026

    Panorama of Entrepreneurship & Career

    16th edition, 27–29 March 2026, Athens Concert Hall (Megaron). Greece’s largest career-advice gathering.

    pan-orama.org →
  • 2026

    Digital World Summit Greece

    Greece’s national initiative on the democratic governance of new technologies. 3 December 2026, French Institute of Athens. Theme: Tech & AI Sovereignty. Under the auspices of the United Nations, the Ministry of Digital Governance and the City of Athens.

    digitalworldsummit.gr →
  • 2018–2019

    Pharmacy Management & Communication

    Recurring contributor to Greece’s leading pharmaceutical magazine, on operations, retail strategy and the slow craft of a well-run pharmacy.

  • 2016–2017

    Student-Staff Liaison Committee, University of Warwick

    Elected course representative for the MSc in Healthcare Operational Management. Carried the cohort’s voice into the faculty’s curriculum and quality reviews.

Education

Where the discipline was trained.

  • 2017–2018

    National & Kapodistrian University of Athens

    Programme “Entrepreneurship in Action”, Centre for Innovation and Entrepreneurship, Faculty of Economics.

  • 2016–2017

    University of Warwick

    MSc, Healthcare Operational Management.

  • 2015–2016

    University of Piraeus

    Diploma in Community Pharmacy Management.

  • 2014–2016

    University of Winchester

    MBA. 50% scholarship from the National Pharmaceutical Association.

  • 2009–2014

    Comenius University, Bratislava

    MPharm, Faculty of Pharmacy.

  • 2008–2009

    University of Crete

    Biological Sciences (foundation year).

Across industries

Life sciences. Financial services. Industrials. Across Europe, North America, Japan and South America.

The actual skill is moving between deeply technical teams and senior stakeholders without losing either of them.

Efthymios Vogiatzis on a manufacturing floor

Writing

Recent essays.

June 2026 · Essay

AI in drug development: a pharmacist's view on data integrity

Where AI actually moves discovery, where data drift quietly poisons it, and what a pharmacist's eye catches that a dashboard does not.

Read

March 2026 · LinkedIn

Hiring the right software

On choosing tools the way you'd choose a colleague, for judgement under pressure, not for the demo.

Read

March 2026 · LinkedIn

Generative AI at work

Where it actually changes the day, where it just adds a tab, notes from inside large enterprises.

Read

A closing thought

The demo is cheap.
The decade is not.

Every advisor can show you a quarter. Fewer can sit with you through the ten years where the decision either compounds or quietly costs you. That is the line VE1 was built on.

Pull up a chair ☕

If your systems do not quite agree with each other, and everyone has been too polite to say so, you will feel at home here.

efthymiosvogiatzis@gmail.comLinkedIn