Get leads, not likes | GM @ Taplio | Founder @ Anser | ex @Ledger | Relentlessly building in Marketing and Tech

I’ve spent over 20 years building and scaling teams across product, growth, tech and operations.

Today, I lead the Social Selling Platforms BU at lemlist (Taplio and Tweet Hunter), where I focus on turning startup momentum into repeatable, scalable systems.

Over the years, I’ve helped founders structure leadership teams, create alignment between product and growth, and scale operations without killing velocity.

I occasionally advise early-stage companies making that leap from startup to scaleup—particularly when growth outpaces structure.

If you’re building a fast-growing company and need perspective on scaling people, process, or focus, feel free to reach out.

Patrice Orenes-Lerma

Stop asking employees to “post on LinkedIn”.

It never works.

They’re busy.
They don’t know what to say.
And the moment you push too hard, it feels fake.

So here’s a better way:

Launch employee advocacy in 30 days.

With 3 to 5 employees.

And without asking for anyone’s LinkedIn password.

The goal is simple:

Stop using the company page as a megaphone.

Start using employees as the distribution engine.

This guide breaks it down into 4 phases:

Days 1 to 7
Pick 3 to 5 employees and set a secure ghostwriting workflow.

Days 8 to 14
Build a month of content that sounds like them, without stealing their time.

Days 15 to 21
Go live and win distribution with a daily 10 minute engagement routine.

Days 22 to 30
Double down on what works, repurpose winners, and scale to more employees.

If you’re stuck at “we should do employee advocacy” but it never ships...

This is the playbook.

Want the guide?

Comment “launchpad” and I’ll send it.

General Manager @ Taplio / Tweet Hunter | Helping startups scale into structured, high-performing organizations

Most companies treat LinkedIn like a loudspeaker.

Blast announcements.
Hope for reach.
Wonder why it's dead on arrival.

Here's the problem: nobody cares what a brand says. 

People care what people say.

But forcing employees to post won't fix it.
They're not marketers. They have better things to do.
Push too hard, and it backfires. Feels forced, turns fake, nobody wins.

So, skip the ask. Build a small squad. Pick a few who actually want in. Give them real support, not a pile of "brand guidelines."

Set up ghostwriting that feels like them, not HR. Make content creation fast, painless, invisible. No passwords. No policing.

Let them own their voice. Run a tight, simple process, batch what works, ditch what doesn't. Scale only after it clicks.

The result: humans speak, people listen, reach grows.

Brands don't go viral. People do.

🤫 comment "secret" and I'll send you the secret 30 day employee advicacy launchpad everyone is looking for

No alternative text description for this image

This week I spoke with a user who made $500k from LinkedIn with 2 000 followers.
The method is stupidly simple.

No audience.
No virality.
No personal brand.

This works for one reason.

He writes for one buyer.

A B2B founder selling a high-ticket service.
Long sales cycle.
Quiet buyers.

Every post follows the same rules.

One post.
One person.
One problem.

No teaching.
No frameworks.
No thought leadership.

Just situations the buyer already lives.

Before publishing, there's only one question:
Would this person want to talk to me after reading this?

If not, it doesn't go out.

Most posts get 10–15 likes.

That’s the signal.

The right people don't engage.
They save.
They come back weeks later.

Then they DM.

To put this in place and stay consistent, he used the Stupidly Simple Content System.

That’s how $500k happens with 2k followers.

Comment "stupid" and I'll send you the Stupidly Simple Content System !

P.S.: Bonus, a picture of me because I can't wait to be on the slopes

Everyone says they want to win on LinkedIn.

Almost nobody sticks around long enough for anyone to notice.

Not because they're lazy. Not because LinkedIn is stacked against them.

They just build these over-complicated systems that collapse the second real work or life gets in the way.

Week one, they're possessed. 
Posting every day. 
Obsessing over what time works best. Chasing some mythical algorithm secret that lives in a podcast from 2021.

Week two, reality hits. Work gets busy. Ideas disappear. Suddenly, LinkedIn feels like another job with no upside.

By week three, they're ghosts.

Here's the truth nobody wants to admit.

Discipline isn't the problem. Complexity is.

If your system takes more than five minutes to explain, it's not a system. It's a ticking time bomb.

I've watched dozens of smart, ambitious people repeat this pattern. They try to outsmart LinkedIn. They act like it's a hackathon, not a habit.

All the clever tricks in the world don't matter if you can't stick with it.

The survivors? The ones who actually get the compound benefits?

Their process is deadly boring.

Three types of posts. One slot in the diary. Dump a pile of ideas in a simple doc. Five minutes to scan notifications, respond, log out.

No dashboards. No hacks. No ritual sacrifices to the gods of engagement.

If your process makes you feel clever, it's doomed.

If your process is boring and predictable, you might just stick with it long enough for your name to mean something.

Most people confuse activity with impact. They burn out chasing vanity metrics, then disappear before the real results even show up.

LinkedIn rewards the people who outlast the noise.

Simple always wins.

If you want to actually be remembered, cut the drama and get boring.

Still convinced your fifteen step colour-coded content planner will make you famous? Good luck.

Want to actually get results?

Start showing up.

Keep it simple.

Comment "simple" and I'll give you a system that actually works, for free !

Most people still treat exec posts like gospel.

Hot takes from the top. Strategy memos dressed up as insights.

But nobody buys it anymore. Not candidates, not customers, not even other execs.

People want blood and sweat. Not another CEO rant.

Show me the engineer who broke production at 2 AM and shipped the fix before sunrise.

Show me the marketer who admits which campaign flopped and which one actually made money.

Show me the sales rep who rewrote their own script after losing a deal.

The trust shift is brutal. Authority isn't handed down from the corner office. It's earned in the comments, in the details, from the people doing the work.

Companies still stuck on leadership branding are sleepwalking into irrelevance.

The loudest, most credible voice in your org is never the CEO. It's the team that knows where the bodies are buried.

You want attention in 2026? It won't come from another executive LinkedIn post.

It comes from proof, not polish.

Do you want to start having your employees posting and not just liking ? 

Comment "playbook" and I'll share with you a free 30 day playbook to launch your employee advocacy program!

Everyone thinks the answer is to just post more.

It isn't.

If you want LinkedIn to actually move numbers, stop treating it like a dumping ground for content.

Posting and hoping is not a distribution strategy.

Most companies crank out updates, blogs, or "insights" and wonder why nobody bites.

Here's the brutal truth: nobody cares about your fifth product update this week.

You want a real distribution channel? Make your people part of the engine.

- If your employees aren't engaged, LinkedIn is just noise
- If your leaders never comment, you're invisible
- If your team never amplifies, your content dies on arrival

You don't need fancier graphics or an AI to write punchier posts.

You need buy-in.
You need repeatable systems.
You need actual coordination.

Distribution happens when everyone rows the same direction.

I've seen teams with average content crush metrics, because their people show up, comment, share, and actually talk to prospects in the open.

I've seen "content machines" wither because every post is a solo act.

It's not about more content.
It's about more company.

Want LinkedIn to work for you?
Start by getting five coworkers to comment on your next post.
Watch what happens.

Try it.
Report back.
Tag the silent lurkers who owe you a comment. 

Let's see if your "distribution" is real or just a pipe dream.

You're measuring employee advocacy success by "number of posts".

That's like measuring a restaurant by how many plates they own.

Looks busy. Means nothing.

Volume is the easiest metric to game, and the least likely to move the needle.

Most teams crank out content, log the numbers, then wonder why nothing changes.

It doesn't matter if the company LinkedIn Slack channel is full of daily copy-paste links. If nobody cares, you're just filling empty plates.

Real advocacy leaves a mark.

Not everyone needs to post. Not every post needs to go viral.

But every time someone shares, you should ask:

Did this start a real conversation?
Did anyone outside the company notice?
Did it make someone think, reply, or want to learn more?
Is the employee proud to have their name on it?

Skip the vanity metrics.

Track what matters: reach, engagement, replies, questions, DMs, sentiment.

If you're still counting posts, you're not measuring advocacy.

You're counting plates in the kitchen while customers walk out the door.

Raise your standards or stay irrelevant.

Overcomplicating LinkedIn is the fastest way to disappear.
People show up with a 12-step system, a Notion dashboard, and a spreadsheet for tracking "engagement velocity."

Week one, they're everywhere.
Week two, they're exhausted.
Week three, they're gone.

It's not a discipline issue.
Nobody can keep up with a process that feels like a second job.

LinkedIn isn't won by the cleverest workflow or the fanciest automation.
It's won by the people who can show up, again and again, without burning out.

Here's the play:
Pick a format.
Pick a slot.
Make it so simple you could do it with your eyes closed.

Forget "best time to post" hacks.
Forget obsessing over reach graphs and content buckets.
If you need a manual for your own process, it won't last.

Boring works.
Predictable wins.

The people you remember on LinkedIn?
They're not the ones with the most creative setup.
They're the ones who kept going after everyone else flamed out.

You don't need to optimize.
You need to endure.

Complexity is just procrastination in disguise.

If your posting routine is more complicated than brushing your teeth, it's doomed.

Simple scales.
Repetition compounds.

The rest is noise.

Comment "simple" and I'll send you the simplest system to grow on LinkedIn.
For free

If you use your sales brochure as the base for LinkedIn posts, you’re setting yourself up to fail.

Not because the offer is bad.

Because the format is wrong.

A sales brochure is made to convince.
A LinkedIn post is made to earn attention.

A brochure starts with:
What we do
How we do it
Why we’re different

A good post starts with:
A situation
A mistake
A sharp opinion
A story people recognize

That’s why brochure-based posts flop.

They feel like a pitch from the first line.

They’re full of words that sound nice but say nothing:
-> Seamless
-> End-to-end
-> Best-in-class
-> Game-changing

On a website, that’s fine.

On LinkedIn, those words trigger one reaction:
Scroll.

It gets worse when you run the brochure through an AI writing tool.

Because AI doesn’t create a point of view.
It amplifies the input.

So it takes generic copy and makes it:
- More polished
- More corporate
- More bland

Which makes the post even easier to ignore.

The fix is simple.

Don’t start from the brochure.

Start from raw material:
-> A real problem you see every week
-> A mistake customers keep making
-> A counterintuitive lesson
-> A moment from a call
-> A belief you’re willing to defend

Then turn that into one idea.

If the post could be copy-pasted onto your homepage, it’s not a LinkedIn post.

It’s a brochure with line breaks.

Struggling to find a method that works for your content creation? Comment "playbook" and I'll share the simplest, faster content system.

I build AI products from concept to production with high-velocity teams. Most recently, I led 7 AI projects at SAP using startup-style squads of 12–25 engineers, shipping each MVP in one quarter or less and delivering ~$65M+ in combined annual business impact.The project I’m proudest of: a RAG-powered chatbot I built hands-on with Claude Code, leading a 12-person team from zero to production in 4 months. It replaced Tier-1 support entirely, cut escalations 40%, and saved $19M/year. I was 50% hands-on, architecting the retrieval pipeline, writing code, and debugging alongside my engineers.What I do:• Scale engineering organizations: Grew teams from 12 to 100+ across 4 continents (NA, India, Brazil, Prague), managing $25M P&amp;L with 99.999% uptime across platforms serving 60M+ users• Ship AI at enterprise scale: 7 production AI systems, including NLP sentiment analysis, OCR mobile apps, real-time AI auditing, ML risk scoring, Recruiting Copilot, Skills Graph, and Supplier Risk platforms• Operate as a player-coach: I still architect systems, review code, and pair-program with my senior engineers. I believe the best engineering leaders stay close to the code.Domains I’ve built in: Enterprise SaaS, Healthcare (radiation oncology, DNA sequencing, cardiac devices - FDA/ISO 13485), E-commerce, FinTech, AI InfrastructureStack: GenAI/LLMs, RAG pipelines, Agentic AI workflows, Kubernetes, AWS/Azure/GCP, distributed systems, microservices, CI/CD, MLOps, observabilityCompanies: SAP (9 yrs), Oracle, IBM, Macy’s, Broadcom, Varian Medical, Thermo Fisher, St. Jude MedicalI write about AI engineering leadership and building high-velocity teams. Follow along if you’re scaling an AI org or hiring engineering leaders who ship.

Debasish Bhattacharjee

I watched a junior engineer build in one afternoon what used to take my team two weeks.

And my first reaction was not pride. It was fear.

He was using an agentic coding tool to scaffold an entire API integration. The kind of work that, when I was a senior engineer fifteen years ago, required three people and a two-week sprint.

He finished before lunch. And the code was cleaner than what most of my mid-level engineers write manually.

I sat in my home office staring at the pull request and felt something I had not expected. Threatened. For about twenty minutes, I genuinely wondered if everything I had learned over two decades was about to become irrelevant.

Then I reviewed the code more carefully.

The architecture was solid. The syntax was perfect. But the error handling assumed a happy path that would collapse under real production traffic. There was no circuit breaker. No retry logic. No graceful degradation for the 14 downstream services that depend on that API.

The tool wrote code. It did not write experience.

I spent an hour with him walking through what happens when 50,000 concurrent users hit an endpoint designed for 5,000. What happens when the third-party service returns a 504 after your retry budget is exhausted? What happens when the database connection pool is saturated at 3 AM, and the on-call engineer has 9 minutes to figure out why?

He looked at me and said, "None of this was in the prompt."

Exactly.

The engineers who will thrive are not the ones who can write code the fastest. They are the ones who know which questions the tool will never think to ask.

AI writes the code. Experience writes the guardrails.

VP Engineering | Scaling AI/ML Organizations from 0-to-Production | 100+ Engineers | $25M P&L | GenAI · Agentic AI · Platform Engineering

I watched a senior engineer present an architecture proposal last week.

Thirteen slides. Clean diagrams. Everyone in the room was nodding.

I asked one question. "What happens to the message queue when the payment service goes down during a peak window?"

The room went quiet.

He looked at me and said, "That shouldn't happen."

I said, "It happened at a company I worked at. A Thursday afternoon. Took us eight hours to recover. We lost forty-seven thousand transactions."

I saw the shift on his face. Not anger. Embarrassment. The kind that comes when you realize you missed something that should have been obvious.

I felt guilty watching it. Because I have been him.

Five years into management, I stopped attending architecture reviews. Stopped reading pull requests. Told myself I was "focusing on strategy." My technical director pulled me aside one afternoon in Palo Alto and said, "You approved a design last week with a single point of failure in the payment path. Did you catch it?"

I hadn't. The shame from that conversation rewired how I lead.

This is what gets misunderstood about senior engineering leadership. People assume that once you reach a certain level, your job becomes decks and dashboards and quarterly reviews.

Some leaders make that trade. I've watched it happen. A VP stops attending architecture sessions. Starts delegating every technical call. Within eighteen months, they can't tell the difference between a good proposal and a polished one.

Their teams notice first. Across my teams spanning multiple continents, engineers test you constantly. Not maliciously. They bring a problem to a meeting and watch whether you ask surface questions or the ones that cut three layers deep. That's how they decide whether to trust your judgment or just comply with your decisions.

The skip-levels go quiet. Hard questions stop surfacing. Not because the problems went away, but because everyone knows the answer will be "let's take that offline."

Last quarter, staying technical caught three critical issues before they hit production. A database migration that would have locked a table serving thirteen million daily queries. An API contract change that broke backward compatibility for 340 consumers downstream. A logging pipeline is silently dropping nearly a quarter of error traces during peak traffic.

None showed up in dashboards. None were in status reports. They surfaced because I asked specific, uncomfortable questions that only come from years of building and breaking systems yourself.

Technical credibility is not a phase of your career you eventually outgrow.

It's the foundation that makes every other leadership skill worth anything.

undefined undefined

How to Use Generative AI: Building an AI-First Mindset

Positioning and Selling SAP Cloud ERP and RISE with SAP Solutions - Record of Achievement

Exploring the SAP Business AI Portfolio - Record of Achievement

Exploring SAP’s Cloud Strategy - Record of Achievement

Betty Liu on Career Success

SAP is the leading enterprise application and business AI company. We stand at the intersection of business and technology, where our innovations are designed to directly address real business challenges and produce real-world impacts.

Our solutions are the backbone for the world’s most complex and demanding processes. SAP’s integrated portfolio unites the elements of modern organizations — from workforce and financials to customers and supply chains — into a unified ecosystem that drives progress.

SAP privacy statement for followers: www.sap.com/sps

Our Community Guidelines
At SAP, we're committed to fostering meaningful conversations that respect everyone in our community. To maintain a positive environment, we moderate comments that:
• Target individuals personally, including our employees, customers, or partners
• Contain discriminatory, harassing, or threatening language/content
• Share personal information without consent
• Promote misinformation or spam or 3rd-party links

We believe in open dialogue and constructive feedback, but we will remove content that violates these guidelines without notice. We appreciate your understanding and contribution to a respectful community.

For questions about our moderation practices, please DM or contact us at brandsocial@sap.com.


Led 7 AI product launches using startup-style execution: assembling 12–25 person cross-functional squads, shipping MVPs in one quarter, and scaling to production across platforms serving 60M+ users with 99.999% uptime.

Reported to VP Engineering. Managed $25M P&L across 100+ engineers on 4 continents.

Key AI Projects Delivered (each 0-to-production):

• RAG Chatbot (Claude Code): Built hands-on with a 12-person team. Replaced Tier-1 support, reduced escalations 40%, saving $19M/year. 50% hands-on, architecting retrieval pipelines, writing code, and debugging in production.

• ExpenseIt AI OCR Mobile App: Led ML-powered receipt scanning for iOS/Android. Achieved Apple App Store and Google Play’s highest satisfaction ratings for the category.

• Concur Detect Real-Time AI Auditing: Shipped NLP-based anomaly detection for expense fraud. $9.5M annual savings.

• Concur Embedded ML Risk Scoring: Deployed predictive risk models across the Concur platform. $8.8M annual savings.

• SuccessFactors Joule Recruiting Copilot: Built a GenAI-powered recruiting assistant. $9.8M annual savings.

• Talent Intelligence & Skills Graph: Designed an ML skills ontology system mapping employee capabilities. $8.7M annual savings.

• Ariba Guided Buying & Supplier Risk: Shipped supplier risk ML models and guided procurement AI. $9.9M annual savings.

Director of Engineering & Operations, AI/ML Platforms

Owned end-to-end cloud and mobile platform delivery across web, iOS, Android, Mac, and 25+ device/OS combinations for SAP Concur and TripIt, platforms used by 60M+ travelers globally.

• Relaunched TripIt in 6 months with a 25-member team: migrated to AWS, redesigned UX, boosted traffic 22%, regained 5M users, and grew revenue 42%.

• Cut manual testing 60% through vendor partnership with Testlio and Sofy, boosting release readiness 35% in 2 quarters and saving $1.2M/year in QA costs.

• Achieved 99.999% SLA in 1 year for 70M+ API requests by optimizing CI/CD and backend, raising velocity 35% and defect detection 30%.

• Led engineering across 25+ devices and 4 platforms, improving system testability and UX validation speed 70%, delivering GDPR-compliant mobile workflows in 6 months.

Senior Engineering Manager, Cloud & Mobile Platforms

Built and scaled SAP Concur’s platform engineering and infrastructure automation from the ground up. Grew team from 5 to 8 engineers across 5 scrum teams spanning mobile and web infrastructure.

• Led 8 engineers across iOS, Android, and web infra: cut delivery delays 30%, improved release consistency 40%, saved $850K in 2 quarters.

• Scaled cross-functional hiring pipeline 50%, drove 35% delivery consistency improvements, enabled $1.2M/year cost savings via systemized infrastructure upgrades.

• Engineered CI automation via Jira/Zephyr, cut infra backlog 80%, sped up stability tests 50%, and hit compliance benchmarks across mobile pipelines in 6 months.

• Unified test automation framework across Browser, Mobile (iOS & Android), and Microservices — replacing 3 separate frameworks into one, reducing development effort and error triaging significantly.

• Established a fully automated CD/CI test/release pipeline (Apr 2019), enabling continuous delivery.

Engineering Manager, Platform Engineering

•	Engineered CI automation via Jira/Zephyr, cut infra backlog by 80%, sped up stability tests by 50%, and hit compliance benchmarks across mobile pipelines in 6 months.

Senior Developer, Platform Engineering & Infra Automation

•	Hired and led 18 SDETs to automate BRM Cloud deployments using Docker/Chef, cutting validation time by 50% and accelerating CI/CD across Oracle Public/Private Cloud.
•	Delivered secure, DR/HA-ready deployments on Oracle Cloud and BMaaS, reducing incident recovery time by 40%, resolving 95% of Qualys issues, and authoring audit-ready privacy SOPs.

Software Architect

•	Drove platform validation frameworks covering 25+ devices and 10+ locales, reducing manual effort by 60% and accelerating global eComm regression testing cycles.
•	Managed 12 engineers and vendors; defined KPIs and dashboards to track test health, boosting BOPS release confidence by 45% across 6 eComm modules.
•	Migrated core systems to microservices & Kubernetes, cutting deploy time 40% and scaling retail ops; mentored 10+ to deliver real-time Spark insights.
•	Led GCP Retail Search integration (Java/SpringBoot), reducing search drop-offs by 20% and boosting conversion by 15%; cut data latency by 35%.

Engineering Manager

•	Built Clarity SaaS automation for provisioning, upgrading, and VM restore; reduced overhead by
$2M/year, enabled 40+ CI runs/week, cut turnaround 45%, boosted renewals 20%.

Principal QA Engineer

Provided software engineering and systems consulting to leading medical device and healthcare technology companies:

• St. Jude Medical (Sunnyvale) - Cardiac rhythm device systems and integration testing
• Varian Medical Systems - Radiation Oncology treatment platforms development and testing
• Applied Bio systems (CurrentlyThermo-Fisher Scientific) - DNA sequencing and fragment analysis applications automation testing
• Bio-Rad Laboratories - Diagnostic systems
• Apollo Group - Enterprise education technology


Senior Software Engineer Consulting

Developed test automation frameworks for 4 products within 3 years, established software engineeringpractices reducing bugs 65% (150→52/release)

Senior Software Development Engineer

I woke up to 43 Slack messages and knew something had gone wrong.

It was a Tuesday morning in Seattle. 6:47 AM. My phone screen was a wall of red notification badges.

Our team in Prague had hit a critical blocker on a deployment that was supposed to go live by the end of their business day. They needed a decision from me. The decision required context from two engineers in Bangalore who had already logged off for the night.

For sixteen hours, the Prague team waited. Pinged Slack. Waited more. Escalated to a channel I had muted three weeks earlier because it was too noisy.

By the time I read the messages, the deployment window had closed. The client was told we'd need another two sprints to deliver. That was the moment I realized I wasn't managing a global team. I was managing three separate teams that happened to share a Jira board.

I felt something worse than anger. I felt responsible for a failure I could have prevented with one Slack reply at 11 PM.

My manager asked during our next sync, "What went wrong?"

I gave him the honest answer. "I did. I designed the communication flow around my schedule, not theirs."

That was the blindness. I had built every decision tree, every escalation path, every standup cadence around Pacific time. Twenty-three engineers across three continents, and the operating rhythm was built for one of them. Mine.

It took me ten weeks to rebuild the communication architecture. Asynchronous decision documents replaced real-time Slack threads. A rotating "on-call decision maker" across time zones meant no single region waited more than three hours for a critical answer. We created a 15-minute recorded daily handoff between each region instead of one standup that worked for nobody.

The results showed up faster than I expected. Deployment velocity increased 35% within the first quarter. Cross-region escalations dropped from twelve per week to two. And the thing that surprised me most, the Prague team started volunteering for stretch projects they had previously avoided. They told me later it was the first time they felt like full participants, not a satellite office waiting for headquarters to wake up.

I manage teams across four continents. The complexity has grown. But the principle hasn't changed since that Tuesday morning with 43 unread messages.

If your global team is only productive when you're online, you haven't built a global team.

You've built a bottleneck with your name on it.

I got a call from our CFO asking me to explain agentic coding to the board. He gave me fifteen minutes and said, "Make it make sense to people who have never written a line of code."

That was the hardest presentation of my career.

Not because the technology is complicated. Because the business implications are so large that underexplaining it sounds dismissive and overexplaining it sounds like science fiction.

I had watched AI coding tools evolve from autocomplete suggestions to fully autonomous agents that can plan, write, test, and deploy code with minimal human oversight. I had been implementing these tools across my engineering organization of 100 engineers for eight months.

But translating "this agent wrote 340 lines of production code in four minutes" into language a board of directors can evaluate is a different skill entirely.

Here is what I told them.

AI coding is not one thing. It is a spectrum. On one end, you have tools that suggest the next line of code while an engineer types. On the other end, you have autonomous agents that can take a business requirement, break it into technical tasks, write the code, run the tests, and open a pull request for human review.

Most companies are at step one. The ones that will pull ahead in the next 24 months are the ones that figure out step three.

Our board chair asked me a question I was not prepared for. She said, "If AI can write the code, why do we need 100 engineers?"

I told her the same thing I tell every executive who asks that question. "Because writing code was never the hard part. Deciding what to build, why it matters, and how it fits into a system that serves real customers under real constraints, that is the hard part. And that is getting harder, not easier."

She nodded. The CFO approved our expanded AI tooling budget the following week.

If you cannot explain AI to someone who does not build software, you do not understand it well enough yourself.

I interviewed a candidate for a senior engineering role who had eight cloud certifications and could not explain how to deploy an AI model into an existing enterprise pipeline.

That conversation changed how I evaluate talent.

He rattled off every acronym. Azure Foundry. Copilot Studio. Power Platform. He had passed every exam. His resume was a wall of badges.

I asked him one question. "Walk me through how you would take a fine-tuned model and integrate it into a legacy claims processing system that handles 30,000 transactions daily."

Silence.

Not because he was nervous. Because nobody had ever asked him to connect the certification knowledge to a real workflow.

I have seen this pattern accelerate over the past two years. The skill gap in AI is not between people who understand AI and people who do not. It is between people who understand AI in isolation and people who understand AI inside an enterprise.

Certifications teach you what each tool does. Experience teaches you why three of those tools should not be used together in a regulated environment.

I manage engineering teams across four continents. When I look at what my strongest architects have in common, it is not the number of certifications. It is the number of production incidents they have survived.

My head of platform engineering in Dublin told me something that stuck. She said, "I would rather hire someone who has debugged one failed AI deployment than someone who has passed ten exams about successful ones."

She was not being dismissive. She was being honest about what enterprise AI actually requires. Patience. Context. The ability to translate between what the technology can do and what the business needs it to do inside a system that was built fifteen years ago.

The gap is not technical knowledge. The gap is integration wisdom. And no exam tests for that.

I still draw architecture diagrams by hand. And it might be the most valuable habit I have as a Director managing 100+ engineers.

A few months ago, I was reviewing a system design proposal from one of my senior architects in Bangalore. Thirty slides. Beautiful. Color-coded. Every microservice is labeled. Every API endpoint is documented.

I could not follow it.

Not because it was bad. Because it was so polished that the complexity was hidden behind the aesthetics. The diagram looked clean, but the system it described had 47 points of failure and no clear fallback path.

I opened a blank canvas in my editor and started sketching. Boxes. Lines. Arrows. No formatting. No colors. Just logic.

Fifteen minutes later, I sent the sketch to the team. My architect looked at it and said, "Oh. I see the problem now."

The problem was a circular dependency between three services that the polished diagram had buried inside a single labeled box.

I felt a little embarrassed at first. Here I was, a Director with two decades of experience, drawing boxes like a junior developer. But that sketch probably saved us four months of debugging in production.

There is a reason I still insist on diagramming my own systems. When you draw something yourself, you cannot hide from what you do not understand. Every gap in your knowledge shows up as a line you cannot connect.

I have watched engineering leaders slowly stop doing this. They delegate the thinking along with the drawing. And then they sit in architecture reviews nodding at diagrams they cannot actually read.

Tools change. The diagramming habit does not.

The leaders who can still sketch a system on a whiteboard in real time are the ones who catch the bugs that survive every code review, every test suite, and every staging environment.

Stay close to the work. The moment you stop drawing, you stop seeing.

I had 130 engineers spread across seven time zones, and I couldn't figure out why my best ideas were coming from the same twelve people.

Dashboards were green. Sprints completed on time. Velocity has been trending three straight quarters upward.

I told myself the system was working.

Then I flew to Prague for a quarterly review. After the session, an engineer named Tomáš caught me in the hallway. He just waited until no one else was around.

He said, "We build what you tell us to build. But nobody asks us what we think should be built."

I asked how long he'd felt that way.

"Since the third month."

We were sixteen months in.

I felt ashamed. Not the loud kind. The slow kind that creeps in when you realize you've been the problem you keep trying to solve.

But here's the part I don't hear other leaders admit. I already knew something was off. I'd seen it in sprint retros. Short answers. No pushback. Cameras off. I told myself it was a cultural difference. That some teams are just quieter.

That's the lie leaders tell themselves when fixing the real problem means giving up control.

Centralized planning is easier to manage. One roadmap. One time zone makes decisions, and everyone else executes. It's clean. It's efficient. And it slowly kills every original thought outside your headquarters.

I didn't fix it because Tomáš opened my eyes. I fixed it because he said out loud what I'd been choosing to ignore.

For over a year, engineers were joining planning sessions at midnight their time, receiving decisions already shaped before they logged on, raising hands into rooms where no one was watching.

I had turned thinkers into ticket-takers. My dashboards never caught it because compliance looks identical to commitment on a burndown chart.

I restructured everything in eleven weeks. Killed centralized planning. Gave each region architecture ownership and roadmap sessions during their peak hours, not ours.

The hardest part was watching engineers hesitate to speak three weeks into the rollout. Trust doesn't rebuild on a new meeting invite. It rebuilds when someone speaks up, and nothing bad happens. Then again. Then again.

Two quarters later, engagement for distributed teams climbed from 54% to 83%. Attrition dropped from 28% to under 9%. One remote team shipped its first independently proposed feature within four months.

Tomáš stayed. Got promoted. Leads a team of eight.

But what I think about most isn't the metrics. It's how long I knew and didn't act. How many retros do I sit through where silence should have been a fire alarm, and I called it a clean meeting?

Most leaders with distributed teams know something is broken. They feel it in the short answers and the cameras that stay off. They just don't fix it because the current structure is easier to run.

Your org chart says "global team." Your calendar says who actually matters.

I built a RAG chatbot that replaced our entire Tier-1 support operation.

12-person team. 4 months from zero to production. $19M/year in savings.

Here’s what nobody tells you about shipping RAG at enterprise scale:

The retrieval pipeline is 80% of the work. The LLM is the easy part.

We spent the first 6 weeks just getting the data layer right - chunking strategies, embedding models, re-ranking logic. I was hands-on writing code alongside my engineers using Claude Code.

What we learned:

1. Chunk size matters more than model choice. We tested 4 different chunking strategies before finding one that didn’t hallucinate on edge cases.

2. Evaluation is harder than building. We built custom eval pipelines before we built the product. If you can’t measure retrieval quality, you’re guessing.

3. The “last mile” is enterprise integration. Connecting to auth, permissions, and audit logging took as long as building the core RAG pipeline.

4. Start with the hardest queries, not the easiest. We deliberately tested failure cases first. This saved us from a false launch.

The chatbot now handles 40% of all support queries that used to go to human agents.

This is what I mean when I say I operate as a player-coach. I don’t just lead teams. I build with them.

If you’re shipping RAG to production, I’d love to compare notes.

#AI #RAG #EngineeringLeadership #GenAI #LLMs

I tried to organize 12 terabytes of data using AI last weekend.

Twenty-five years of life sitting on spinning disks. Family photos from three phones across two decades. Legal documents. Tax returns from 2006. Files I forgot existed.

That was just the external hard drive.

I also had a 4-terabyte flash drive full of dashcam footage from my electric vehicle. Four cameras. Continuous 4K video. And a NAS with years of cluttered backups across multiple RAID volumes.

Three devices. Nearly 24 terabytes of mess.

I read through Claude Cowork's technical docs and watched YouTube vloggers demo it. Cowork is essentially a vibe-coded Claude, not a separate product, and I wanted to experiment with that.

Every single tutorial showed the same thing. Clean your Desktop. Organize your Downloads folder.

That is a four-minute task. Not a real problem.

I started with Claude Cowork on my local machine. One terabyte internal drive. It restructured folders, flagged duplicates, moved them to staging, generated an artifact report, and gave me one prompt to authorize deletion. Even sent a Slack notification with logs when done.

Took about 135 minutes. Solid.

Then I plugged in the external drive. Nothing. Claude Cowork cannot mount external volumes. No tutorial mentions this.

Switched to Claude Code from my terminal. It is estimated to run for 16 to 20 hours interactively. Massive token consumption across 24 terabytes of scanning, renaming, and deduplication.

I pay $200 a month for Claude Code Max. The system allocates daily sessions and has a weekly max limit. At this scale, I would have maxed out my entire weekly allocation within a few hours and waited a full week to resume.

So instead of having AI do the work, I had it write the work.

I spent thirty minutes engineering the prompts. How to reduce token cost. How to structure a feedback loop. How to run evals. How to guarantee every duplicate flagged was a true duplicate, not a false positive. 

I fed Claude Code the constraints, the validation requirements, and the cost-reduction targets. 

Then it built a sophisticated Python script with internal checks, iterative error correction, and eval logic baked in so the output would not fail at scale. MD5 checksums. Folder restructuring. Duplicate staging. Video compression. Automated permissions. All running locally after the script was complete.

Building that script cost roughly 9.63% of my weekly token allocation. The remaining 90.37% stayed untouched. Once the script was running locally, zero tokens were consumed.

The script ran across all three devices in about eight hours. Found 11,573 duplicates. Freed 2.3 terabytes. Compressed 340 gigs of dashcam footage to 87 gigs.

Every AI demo online solves a clean, laptop-sized problem. Nobody shows you the real ones sitting in desk drawers and closets.

Sometimes the best use of AI is not asking it to do the work. It is asking it to build the thing that does the work.

How I Used Claude Code to Organize 24TB of Data

If you want to kill an AI project, make sure nobody owns it.

I watched it happen with a project I was responsible for. Not from a distance. I was the one who set it up to fail.

We had a machine learning initiative in Manila. Twelve engineers. Seven months of runway. Executive sponsorship. Budget approved. The goal was straightforward. Build an ML model to predict customer churn in a SaaS platform and integrate it into the existing workflow.

Data science built the model. Platform engineering was supposed to build the serving infrastructure. Product management defined the requirements. QA was responsible for validation in a regulated environment.

On paper, everyone had a role. In practice, nobody owned the space between roles.

The model was ready in eight weeks. It sat in a notebook for five more months because nobody owned the pipeline that would move it from experiment to production. Data science said it was an infrastructure problem. Platform engineering said they were never told the model was ready. The product blamed engineering. Engineering blamed the product.

My VP called me into his office on a Wednesday afternoon and said, "I just told the board this would be in production by Q3. Where is it?"

I didn't have a good answer. The embarrassment was physical. A tightness in my chest that I can still recall.

Here is what I was blind to. I had structured the project like a traditional software initiative with clear team boundaries. But AI projects don't respect team boundaries. The handoff between data science and production engineering is where every ML initiative either lives or dies. And I had treated that handoff like a line item on a Gantt chart instead of the single most critical risk on the project.

I created a new role on the spot. Technical Program Manager for ML integration. One person whose only job was to own the gap between teams. No feature work. No model training. Just the ugly, unglamorous work of making sure what data science built could actually run where it needed to run.

Within nine weeks, the model was in production. Within six months, it was driving a 28% improvement in early churn detection. The project that had stalled for five months shipped in less than ten weeks once someone owned the middle.

Every AI initiative I've led since then starts with the same question. Not "who builds the model" or "who owns the infrastructure."

Who owns the space between?

That's where AI projects go to die. Or to ship.

I became an early OpenAI beta customer in 2022 when most engineering leaders I knew were still calling AI a science project.

I remember the exact moment I realized how wrong everyone around me was. I was sitting in a quarterly planning meeting in San Ramon, and our head of product said, "AI is a nice experiment, but our clients are not asking for it."

Six months later, every single one of those clients was asking for it.

What I did not expect was how isolating it would feel to be early. I spent most of 2023 having the same conversation over and over. "Is this real?" "Is this just hype?" "Should we invest now or wait?"

I stopped trying to convince people with slides. I started building prototypes. Small ones. A document classifier that reduced manual review time from 3 hours to 12 minutes. A compliance checker that caught 23 regulatory gaps our legal team had missed across two quarters.

The prototypes did the convincing I never could.

My CTO at the time told me something I think about often. She said, "I did not believe you until I saw it working in production. And honestly, I am a little ashamed it took me that long."

That was not a failure of intelligence. It was a failure of imagination. Most leaders I have worked with do not resist AI because they are skeptical. They resist it because they cannot visualize what it does to their specific workflow until someone shows them.

Three years later, I manage teams across four continents, building AI into enterprise systems. The technology is different. The conversation is different. The budgets are different.

But the pattern is the same. The leaders who adopt AI the earliest are not the most technical. They are the most curious. And curiosity is the one trait no certification can teach.

The best time to start building with AI was in 2022. The second-best time is before your competitor finishes reading this post.

I learned more about engineering from an FDA auditor than from any CTO I've worked for.

He walked into our Dublin office on a Monday morning. Gray blazer. No laptop. Just a legal pad and a pen. He spent nine hours asking questions that none of my eleven engineers could answer to her satisfaction.

Not about code quality. Not about test coverage. About traceability.

"Can you show me the exact commit that changed this calculation? Can you prove that the version running in production is the version that was validated? Can you trace this output back to the requirement it was built from?"

We could answer some of it. Not all of it.

I watched my most senior architect struggle to explain our deployment pipeline to someone who didn't care how elegant it was. He only cared whether it was provable.

At the end of the day, he looked at me across the conference table and said, "Your engineering is strong. Your evidence trail is not."

I felt a specific kind of dread that I hadn't experienced before in twenty-two years of building software. Not the kind from a production outage. The kind of realization that everything you built might not count because you can't prove how you built it.

That audit didn't fail us. But it came close enough that I couldn't sleep properly for two weeks.

Here's what I was blind to. I had spent my career optimizing for speed, scalability, and reliability. In healthcare and regulated environments, none of that matters if you can't demonstrate compliance at every layer. Nobody had ever forced me to think about engineering as evidence.

I rebuilt our CI/CD pipeline with traceability as a first-class requirement. Every commit is linked to a requirement. Every deployment is tied to a validated build artifact. Every test is mapped to a regulatory control. It took seven months. My team resisted for the first three.

Twenty-one months later, we passed our next audit without a single finding. The auditor said, "This is one of the cleanest submissions I've reviewed this quarter."

My architect, the one who'd struggled in that first audit, pulled me aside afterward and said, "I used to think compliance was something legal handled. Now I think it's the hardest engineering problem I've ever worked on."

The companies building healthcare SaaS right now and treating compliance as a checkbox are building on a foundation that will crack the moment someone with a legal pad walks through their door.

Compliance isn't a burden you carry alongside engineering.

It's engineering. Just the kind nobody teaches you in school.

I tracked the wrong metric for fourteen months, and my leadership team celebrated the entire time.

We were measuring model accuracy. Every monthly report showed improvement. Started at 72%, climbed to 91% over the course of a year. Charts went up. Executives nodded. I presented it to the board as proof that our AI investment was working.

Then a customer called.

Not to complain about the AI. To ask why their account manager hadn't followed up on a churn signal that our model had flagged three weeks earlier. The model had predicted the churn correctly. Nobody had acted on it.

I pulled the data that evening in my home office in San Francisco. Sat there until almost midnight, going through seventeen months of model outputs. The accuracy was real. But the adoption rate of model recommendations by the customer success team was 23%. Meaning our 91% accurate model was being ignored more than three-quarters of the time.

I felt sick. Not frustrated. Physically sick. Because I had been standing in front of executives for over a year, telling a story that was technically true and operationally meaningless.

My data science lead asked me the next morning, "Did you know the CS team wasn't using the outputs?"

I said, "I didn't even think to ask."

That was the blindness. I was so proud of the engineering achievement that I never measured whether it changed anything in the real world. Model accuracy was my vanity metric. The one that made my team look good without requiring us to do the harder work of making sure humans actually trusted and used what we built.

I changed how we measure AI success after that. We added three new metrics. Recommendation adoption rate. Time from prediction to human action. And revenue impact per prediction acted upon.

Within eighteen months, adoption climbed from 23% to 74%. Not because the model got better. Because we finally started measuring whether anyone cared about its output.

Every company measuring AI success by model accuracy alone is building a trophy case while the game is still being lost on the field.

The metric that matters isn't how smart your model is.

It's whether anyone's behavior changed because of it.

I became an OpenAI beta customer before most engineering leaders could spell GPT.

That was early 2022. No ChatGPT yet. No hype cycle. No boardroom panic about AI strategy.

Just a quiet API and a gut feeling that this would change everything about how my teams build software.

My manager at the time looked at the invoice and said, "What exactly are we paying for here?"

I didn't have a great answer. Just a conviction that felt borderline reckless.

I remember the exact moment I knew I was right. A Wednesday afternoon in San Francisco. One of my senior engineers had spent nearly a week debugging a production issue across 17 microservices. I fed the error logs and trace data into the API. Twenty minutes later, I had a root cause hypothesis. Not perfect. But close enough that he resolved the issue within the hour.

He pinged me on Slack and said, "How did you find that?"

I told him. He went silent for about half a minute. Then typed, "Can I get access?"

That moment should have been a victory. Instead, I felt a strange kind of shame. Because I realized I had been sitting on this tool for months, treating it like my personal experiment instead of sharing it with the team that needed it most.

That was my blindness. I was so focused on proving the technology to leadership that I forgot the people actually doing the work.

Within that quarter, 19 engineers across my org were using it voluntarily. Not because I mandated it. Because the ones who tried it told the ones who hadn't. Adoption happened through hallway conversations and Slack threads, not strategy decks.

That was the part I almost missed. I was so focused on the technology that I nearly forgot the adoption lesson I had learned nearly twenty years earlier at IBM. Tools don't spread through mandates. They spread through proof.

Over the following year, those 19 engineers contributed to a 38% reduction in mean time to resolution across production incidents. Debug cycles that used to stretch across days compressed into hours. My VP finally stopped asking what we were paying for.

The companies rushing to build AI strategies right now are making the same mistake I almost made. They're treating AI as a top-down initiative when real adoption happens engineer-to-engineer, as proven by proof.

If your team isn't experimenting yet, the problem isn't the technology.

Nobody showed them what it can actually do on a Wednesday afternoon with a real problem.

I sat through my fourth AI strategy webinar that year and realized I was part of the problem.

Same slides. Same consultants. Same promise that AI would transform everything by next quarter.

I was nodding along like everyone else. Taking notes, I would never read. Feeling productive without producing anything.

Then I went back to my desk on a Friday afternoon in Berlin and built a chatbot. A terrible one. It took me three days. The thing hallucinated product names we didn't even make. It confidently recommended a pricing tier we had retired two years earlier.

My team lead walked by my screen, looked at the output, and said, "Is that supposed to do that?"

I said, "Absolutely not."

She laughed. I didn't. Because in those three days of failing, I learned more about large language models than I had in four webinars and eleven whitepapers combined.

I learned what temperature actually controls. Why context windows matter more than model size for enterprise use cases. What happens when you feed an LLM your internal documentation without cleaning it first? The answer is chaos. Expensive, confidently wrong chaos.

Here's the part that made me uncomfortable. I had been leading engineering teams for over two decades. Managing budgets. Setting technical direction for organizations with over a hundred engineers. And I had been delegating my understanding of the single most disruptive technology of my career to architects and consultants.

That was the blindness. Not that I didn't know AI. That I had convinced myself I didn't need to.

I decided that month. Every quarter, I would build something with my own hands using whatever the latest AI tooling offered. Not production code. Prototypes. Ugly, broken, embarrassing prototypes.

Sixteen months later, that practice has changed how I evaluate every AI proposal that crosses my desk. I can smell the ones built on demos instead of real testing. I know which vendor promises are technically possible and which ones require infrastructure my team doesn't have.

My skip-level engineers noticed. One of them told me recently, "You're the first Director I've worked for who actually understands what we're building. Most of them just read the executive summary."

I asked her when she figured that out. She said, "The day you asked about hallucination rates during a vendor demo. Nobody at your level asks that."
The gap between leaders who tinker and leaders who delegate is about to become the most consequential divide in engineering leadership.

The ones who still build will make better decisions.

The ones who stopped will make expensive ones.

taplio

The Battle

Patrice Orenes-Lerma

Debasish Bhattacharjee

Followers

Posts

Engagements

Likes

Performance ScorePatrice Orenes-Lerma's LinkedIn Performance

Performance ScoreDebasish Bhattacharjee's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Patrice Orenes-Lerma

Debasish Bhattacharjee

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Patrice Orenes-Lerma's LinkedIn Performance

Performance ScoreDebasish Bhattacharjee's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScorePatrice Orenes-Lerma's LinkedIn Performance