Marc van Neerven

Independent Software Business Consultant

Fractional CTO 💡 Top Startups Voice⚡ Helping startups make better SaaS

You’ve got vision.

I help you turn it into something that actually ships, scalable, beautiful, and built to last.

As a Fractional CTO, I help founders and leadership teams:

Turn strategy into execution: translating boardroom vision into working code and real traction.

Build lean, high-performing tech teams: in-house, remote, or hybrid, with just the right balance of structure and speed.

Adopt AI the right way: embedding an AI mindset into products, workflows, and decision-making before it’s too late.

Create SaaS that sells itself: through a Product-Led Growth approach that puts value before friction.

Design for humans, build for scale: applying design thinking, solid architecture, and open-web standards to grow sustainably.

My work is transformational: when I leave, your company keeps moving faster, smarter, and with more clarity than before.

I’m not a slide-deck strategist. I still code, design, and build proof-of-concepts myself, focusing (and advocating for) the Open Web, and native cloud to prove what’s possible.

You want your SaaS to grow leaner, smarter, and more human.
Let’s build it.

Startup Expert

Writing blogs about the life of a CTO at Startups and Scaleups. 

Has the world gone mad? 🤯

I just read Paul Hebert's deep dive on the Shadcn radio button ¹.

It’s thoughtful. Detailed. Carefully engineered.

And it perfectly illustrates the problem.

We are writing React components, layering abstractions, wiring state machines, and debating micro-patterns…
…to recreate something the browser does natively (and very well).

What a radio button should be:

<input type="radio" name="beverage" value="coffee" />

That’s it.

Accessible. Keyboard-navigable. Screen-reader compatible. Form-compatible. Zero JS required.

And yet the modern stack says:
- Hide the native control
- Rebuild it with divs and spans
- Reimplement focus logic
- Synchronize state with React
- Hope accessibility survives the refactor

All to achieve… styling.

But styling is not a JavaScript concern.

Styling should come from your Design System, or at least from CSS that lives in a shared asset and is cached separately.

And styling radio buttons has long been extremely easy with pure CSS.

The deeper issue isn’t Shadcn however.
It’s that we’ve normalized rebuilding HTML in JavaScript.

When your design system requires a component layer to render a radio button, something fundamental shifted.

Not everything that looks modern is progress.

Sometimes the most radical move in 2026 is to trust the platform again.

The browser ships a radio button.
Maybe we should start there.

Article in the comments.

¹ Shadcn components have become a sort of automatic goto for LLM agents - together with React and Tailwind 😔...

#puredesignsystem #webstandards #shadcn

Fractional CTO 🚀 AI Mind, SaaS Soul, Human Heart

10x Software Architect/ Engineer, I am embracing with open arms the agentic coding revolution, with my firm I am delivering rapid prototypes in days, and competitive production solutions in few weeks. 

As a Technology lead, my focus is to pick the company vision and translate that into a technology roadmap. I believe that with the right tools, we are capable of amazing things, that's why I am obsessed with supporting startups and creators kickstarting their product with the right engineering mindset since day zero. I specialize in designing scalable serverless solutions (on AWS and Azure), implementing Agentic AI technologies and cross platform mobile apps. Passionate about innovation, my goal is to build efficient, cutting-edge platforms that redefine industries through high-end engineering approaches.

Paolo Cittadini

(Un)Popular opinion:
AWS > Azure


The Azure portal alone is reason enough.
Every click feels like waiting for a PDF to load in 2005. Blade within blade within blade. Meanwhile AWS console just works.
Then there's pricing: 
Need a cheap dev instance? On AWS you spin up a t3.micro and forget about it. On Azure you're doing math trying to figure out which B-series won't bankrupt you.
And permissions? IAM isn't perfect, but it's coherent. Azure RBAC + Entra ID + custom roles + management groups is a complexity tax nobody asked for.
Maybe Azure has its place, especially if you're deep in the Microsoft ecosystem. But for raw builder experience? Not close.
I've shipped production on both. When possible I reach for AWS every time.

Spoiler, my new product is running on Azure 😆 

#aws #azure #cloud #infrastructure

Fractional Tech Lead | Software Engineering Consultant | AWS, Go, Agentic AI, DevOps

I'm building software with a team of 3 engineers.
None of them are human.

Here's how it works:
I break a feature into issues. Each issue gets a dedicated agent team: 
→ Tech Lead (Opus): coordinates, reviews architecture, gates quality 
→ Software Engineer (Opus): writes Go code + CloudFormation 
→ QA Engineer (Sonnet): catches bugs before commit, validates infra

The workflow is fully autonomous. Tech Lead assigns tasks. Engineer implements. QA blocks until code compiles and patterns are correct. Tech Lead routes fixes until everything passes.
Last week I shipped 4 issues this way. Real production code. Lambdas, DynamoDB tables, API integrations.

Where I still step in: 
→ Deployment and runtime debugging 
→ Architectural decisions (service design, cost tradeoffs, general approach) 
→ Anything requiring real world context

The agent team handles the known path. I handle the unknowns.
This is how I build now. Faster. Fewer bugs reaching production. More time for the decisions that actually matter.
If you're curious how this could work for your codebase, DM me.

#ai #softwareengineering #claudecode #golang #buildinginpublic

We spent 15 years mocking waterfall 🤣 

"Too rigid." 
"Nobody knows what they want upfront." 
"Just ship and iterate."

Agile won. Sprints became religion. 
Requirements became "we'll figure it out as we go."
And then AI happened.

Now suddenly we're realizing: if you can't describe what you want with precision, AI gives you garbage. Fast garbage, but garbage.
The bottleneck moved. It's not coding anymore, it's thinking upfront before anyone writes a line.
You know what that sounds like?
Waterfall.

Not the bloated, 200 page spec document version. But the core idea: know what you're building before you build it.
We didn't kill waterfall. We just pretended we didn't need it.
Now it's back wearing a hoodie and calling itself "prompt engineering."

Almost there! Testing voice AI agents is harder than you think.

We're building an AI that handles hotel bookings over the phone. Real guests. Real payments. Real consequences if it breaks.

Testing this manually meant someone calling the agent after every deploy (also via landline, so many 💰 spent). Pretending to be a guest. Going through the entire booking flow. Every. Single. Time.

We are shipping multiple times a day. This was unsustainable.
So we built automated e2e testing for voice conversations:

- Pre-recorded human speech → replayed over real WebRTC and Twilio
- Full pipeline tested: audio → STT → LLM + tools → TTS → playback
- YAML-driven scenarios with assertions and timing
- Barge-in support (interrupting mid-sentence like humans do)
- Dual-channel validation (browser + phone simultaneously)

The technical challenges were wild:

Making a WAV file behave like a live caller requires precise VAD silence patterns. Barge-in needs millisecond timing. Multi-turn responses need event counting. Loudness normalization matters when STT has energy thresholds.

Result: 6-turn booking conversation, 2 channels, fully automated. Check availability → select room → payment details → booking confirmation. All validated against expected transcripts, tool calls, and response content.

Voice AI testing is fundamentally different from API testing. You're validating real-time, multi-modal pipelines where timing and audio fidelity matter as much as logic.

This is how we keep improving the quality of our voice agents with confidence.

There are still a lot of pain point that I would love to bring up, but  I'll save those for next posts.

More on what we're building is coming up next week.

#ai #voiceai #devops #testing #webrtc

Flew home today after a week in Italy.


Visited my hometown. 
Had two real business meetings.
A hotel in Rome wants to pilot our voice agent. A health lab too.
Different industries. Same pain. Not enough people to answer the phone. Too many calls going to voicemail.
This is why I'm building welcomr.
Our agents answer calls, check availability, book appointments, update your calendar, speak multiple languages. All without putting anyone on hold.
No app to install. 
No training needed. 
It just works with what you already use.
Public demo coming soon. 
But if you want to try it now, DM me.

#AI #voiceAI #startups #AgenticAI #engineering

The closed engineering culture we've been building and perpetrating for years is falling apart like a house of cards.

I have to admit, I did not expect it to happen so quickly.
It's fascinating to watch how we went from being a closed lobby of specialists and sarcastic gatekeepers (some of us still are) to just another profession with no security anymore.

What's even more fascinating, in pure 2020s style, is watching this world, my world, split between those living in complete denial and those experiencing it firsthand. 
Whether it's a blessing or a curse depends on which side you're on.
I'm pretty sure the thousands of engineers being left without job security are living it as a curse. 
And NO, not every engineer can quickly land a new role right now.
Let me talk about the Netherlands job market specifically, because that's what I see every day. It's a mess.
The numbers don't lie: 
→ For the first time since 2021, there are more job seekers than job openings → 93 vacancies per 100 unemployed, down from peaks of 130+ 
→ Netherlands ranks 3rd globally in 2026 tech layoffs with 1,700 cuts (ASML alone) 
→ 30,700 tech jobs cut globally in just 6 weeks this year 
→ 55% of companies expect AI to drive layoffs in 2026 
→ UWV received 355 reorganization notifications last year, highest in a decade

What I see on the ground: 
200+ applicants per role. 
Fake job postings everywhere. 
AI selection walls built by the same folks who complained about developers submitting AI-generated resumes two years ago. The irony. Now if you don't use AI to tailor your application, good luck even passing those filters.
Companies don't know what they want. They're looking for nonexistent perfection. They hide behind "we'll get back to you" or simply ghost you. No feedback. No closure.
And here's the painful truth: if a company can leave a role open for 6 months, they probably didn't need that person to begin with.
The market has changed. Yes, some companies still operate the old way. If you're in that bubble, you probably haven't realized what's happening outside of it.

But for the rest of us? Welcome to the new normal.

#engineering #techlayoffs #jobmarket #netherlands #ai

The hidden cost of building your own product isn't what you think.
It's not Claude tokens. It's not cloud bills. It's not your go-to-market strategy or the sales playbook you haven't figured out yet.
The biggest hidden cost is this: you're betting your reputation.
When you build for someone else, you're protected. The company name is on the door. If it fails, you move on. Nobody remembers.
When you build for yourself, your face is on it. Your name. Your LinkedIn. Your network watches.
And then the questions start:
What if it fails?
What if it doesn't work like I promised?
What if people trusted me and I let them down?
Will I still be credible after this?
I think about this every time I demo something to a customer. Every call that goes through our voice agent, every booking it handles, every weird request it manages to solve, that's my reputation on the line. Not some company's. Mine.
This is the weight nobody talks about. The cost that doesn't show up on any spreadsheet.
Every founder I know carries this. The ones who make it aren't the ones without fear. They're the ones who ship anyway.
Putting your face on what you build is terrifying. It's also the only way to build something that matters.

#buildinginpublic #startups #founder #voiceai

WelcomrAI is live!

Few months ago, together with friends, we started building something that scared me.

Not technically. I've been shipping production software for almost 15 years. Go, AWS, distributed systems, the whole thing.
What scared me was putting my name on it. Fully. No client to share the blame. No team to hide behind.

Though, the tech challenge was real. Orchestrating WebRTC, speech to text, text to speech, conversation logic, all in real time with sub-second latency. Layers on layers that need to work together perfectly.

WelcomrAI is a voice agent for anyone that needs constant customer reply. It answers calls, books appointments, speaks multiple languages, runs 24/7.
It's live now. Real calls. Real customers. Real stakes.

It's not perfect yet. But that's the point.
We tailor everything to what the customer actually needs. No generic templates. No "this is how our product works, adapt to it." We adapt to you.

If your business struggles to keep up with incoming calls, if you're missing opportunities because nobody picked up, get in touch. Let's talk.

(link in the first comment)

#buildinginpublic #voiceai #founder #ai #engineering

The skill isn't just prompting (yet), It's knowing what good looks like before the first line is written and encoding that knowledge so your agents know it too.


After my last post about agent teams (swarms), someone asked: "Doesn't this produce spaghetti code?"
Fair question but here's the honest answer: it would, if I let the agents drive.
I don't.
I'm the architect, they're the builders.
Before any agent writes a line of code, I've already decided things like: 
→ The folder structure 
→ The naming conventions 
→ The patterns we follow (and the ones we don't) 
→ How this feature connects to everything else
The secret weapon? SKILLS md

Claude Code lets you define custom skills: markdown files that teach your agents how you work: coding style, design patterns, testing strategie, what libraries to use, what to avoid etc.

My agents don't guess: they read the skills I wrote and follow them.
The agents get a detailed plan, tot a generic: "build me a search endpoint." More like: "build a Lambda that follows the exact pattern in compute.yaml, uses the X package for DynamoDB, returns the same response shape as the other endpoints."

They implement. I review. 
If something drifts from the architecture, it doesn't ship.

Now the truth (at least in March 2026 - it might change in the next months of course 😁 ):
The skill isn't just in prompting (yet), It's knowing what good looks like before the first line is written and encoding that knowledge so your agents know it too.

#ai #softwareengineering #claudecode #cleancode #buildinginpublic

taplio

The Battle

Marc van Neerven

Paolo Cittadini

Followers

Posts

Engagements

Likes

Performance ScoreMarc van Neerven's LinkedIn Performance

Performance ScorePaolo Cittadini's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Marc van Neerven

Paolo Cittadini

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);}Marc van Neerven's LinkedIn Performance

Performance ScorePaolo Cittadini's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreMarc van Neerven's LinkedIn Performance