Founder & Lead ML Engineer @ AI Flow | Ex-Google & Bloomberg | Building AI Agents that scale | 10 years in AI & ML | Building AI Flow, an AI & ML focused consultancy that designs, builds, and scales end-to-end AI products for startups and mid-market companies.


> What we do at AI Flow

- AI consulting - understand how to make most of your data and benefit from the Agentic AI wave.
- AI Strategy & Roadmapping - cut through hype, map the fastest path to ROI
- Custom ML & Gen-AI systems - from data pipelines to fine-tuned LLMs and Agents
- Data platforms - ETL, ingestion pipelines, document understanding end to end solutions
- Embedded AI Teams - senior engineers who plug in and deliver from day 1

> Proof over promises

- 45% pipeline cost & time drop for hyper-realistic video at Metaphysic
- 65% faster ML iterations for Google internal ML teams
- Millions saved automating invoice processing for a top-5 Oil & Gas firm
- Built tools used by 200K daily users, with more than 200 request/second at Bloomberg.
- 85% RAG & LLM pipeline cost drop through data cleaning pipelines for law companies.

> Why clients trust us

“I trusted you to deliver a good result, and that’s exactly what happened.”
- Josua Vieten, CEO @ ExoMatter

“Deep technical expertise across the stack. Mihai understands business context and proposes scalable solutions.”
- John Hack, CPO @ Interflexion

“Mihai makes it easy to get projects done. He is pragmatic, ambitious, and takes initiative.”
Paul Craciunoiu, Founder @ Uplift


> Let’s talk about building the AI future

Open to B2B, remote projects. If you need a sharp, senior team to ship real AI, book a 30-min call: antonmih.ai/meet

Mihai Anton

You wouldn't hire a mechanical engineer who only used AI-generated blueprints.

You wouldn't trust a doctor who only used ChatGPT for diagnosis.

Yet many companies hire "AI engineers" who've never debugged a production system.

Tools change. Fundamentals don't.

undefined undefined

Comment

Watch it here: https://www.youtube.com/watch?v=CLuv-X2Fm_s

Machine Learning by Stanford University on Coursera

Cambridge English Advanced English

Founder & Lead ML Engineer @ AI Flow | Ex-Google & Bloomberg | Building AI systems that scale | 10 years in AI & ML

Founder & Lead ML Engineer @ AI Flow | Ex-Google & Bloomberg | Building AI systems that scale | 10 years in AI & ML | Building AI Flow, an AI & ML focused consultancy that designs, builds, and scales end-to-end AI products for startups and mid-market companies.


> What we do at AI Flow

- AI consulting - understand how to make most of your data and benefit from the AI wave.
- AI Strategy & Roadmapping - cut through hype, map the fastest path to ROI
- Custom ML & Gen-AI systems - from data pipelines to fine-tuned LLMs
- Data platforms - ETL, ingestion pipelines, document understanding end to end solutions
- Embedded AI Teams - senior engineers who plug in and deliver from day 1

> Proof over promises

- 45% pipeline cost & time drop for hyper-realistic video at Metaphysic
- 65% faster ML iterations for Google internal ML teams
- Millions saved automating invoice processing for a top-5 Oil & Gas firm
- Built tools used by 200K daily users, with more than 200 request/second at Bloomberg.
- 85% RAG & LLM pipeline cost drop through data cleaning pipelines for law companies.

> Why clients trust us

“I trusted you to deliver a good result, and that’s exactly what happened.”
- Josua Vieten, CEO @ ExoMatter

“Deep technical expertise across the stack. Mihai understands business context and proposes scalable solutions.”
- John Hack, CPO @ Interflexion

“Mihai makes it easy to get projects done. He is pragmatic, ambitious, and takes initiative.”
Paul Craciunoiu, Founder @ Uplift


> Let’s talk about building the AI future

Open to B2B, remote projects. If you need a sharp, senior team to ship real AI, book a 30-min call: antonmih.ai/meet

AI Flow | We build AI powered products | AI, Machine Learning, and Data managed services

AI Flow

Senior Software Engineer | AI consultant | Working in AI before it was cool

Resume

Created a social network for frequent travelers, where they can post experiences, see what others explore, and follow their steps. Take a look to see how it works!

IdeaList

MihaiAnton has 14 repositories available. Follow their code on GitHub.

MihaiAnton - GitHub

Founder & Lead ML Engineer @ AI Flow | Ex-Google & Bloomberg | Building Agentic AI & ML systems that scale | 10 years in AI & ML

AI Flow #2 is up.

This week: how we helped a client cut spending from $10K/month to $2K/month, without touching the model.

After analyzing the situation, we found the problem was the input.

They were feeding entire HTML pages to the LLM (cookie banners, navigation menus, scripts, SEO content). The model was drowning in noise.

The fix was simpler than they expected. I break down what we did and the results in this week's edition.

Every time a new model drops, I get the same wave of messages from clients: should we rebuild? Should we switch? Are we falling behind?

This week it's GPT-5.4. Next month it'll be something else.

The honest answer is always the same: it depends on what you're optimizing for, and most teams haven't written that down anywhere.

- If cost is the constraint and the new model delivers similar quality for less, run the eval and test it.
- If accuracy drives revenue or directly reduces support volume, measure whether the delta is real enough to justify the migration work.
- If you're upgrading because it's newer and that feels like the right move, that's the one reason I'd push back on.

The things that determine whether your system works didn't change this week. Your data quality didn't change. Your evaluation criteria didn't change. The problem your users need solved didn't change.

A better model sitting on top of a broken pipeline is still a broken pipeline. Test it against the metrics you already track. If they move, ship it. If they don't, look at other areas that actually move them, like the data sources feeding the system.

AI Flow #3 is up.

This week: a client came in ready to invest in AI. Budget approved, timeline set.
We told them: not yet.

After analyzing their setup, we found data in 6 different systems. No tagging, no structure, half of it contradicting the other half.

Building on that foundation would have guaranteed a broken system. So, we made a different plan instead.

"How fast can we see ROI from AI?"

Depends on what you're measuring.


- Time savings? Weeks. We automated proposal generation for a construction company. From 4 hours per proposal to 20 minutes, they felt the ROI immediately.

- Revenue increase? Months. You need time to test, iterate, and prove the feature changes behavior.

- Cost reduction? Somewhere in between.


Most companies expect all three at once, immediately.

My advice: focus on time savings first, as it’s fastest to prove.

"Should we rebuild for the new model?"

The problem isn't the question itself. It's what comes next, as most teams usually ask, "Is it better?" instead of "Is it better for what we need?"

The answer depends on what you're optimizing for. Here are 3 examples.

Should we use the latest model?

2026 isn't about chasing what's new. 

It's about building systems that work regardless of which model is trending.

AI Flow #4 is up. This week: how fast can you actually see ROI from AI?

I’m sharing one case we had with AI Flow: We helped a construction company cut proposal time from 4 hours to 20 minutes; they felt the ROI in week one.

But that's time savings. Revenue increase? Cost reduction? Each has a different timeline.

The problem is that most companies expect all three immediately, and that's why projects stall before showing value.

“Can we automate this?” is one of the most common questions I get.

Sure, we can. But that’s not the first thing you should ask.

Before you jump into workflow scripts or n8n hacks, zoom out:

- Does this process actually need to exist?
- How much time or cost will automation really save?
- Do you have clarity on the pain points you’re trying to remove?

Most teams skip that step. That’s how you end up automating messy processes, which only automate more mess.

"We need AI to analyze our customer feedback."

This request came from a head of product at a 200-person company.

What they really had:

- Feedback in 7 different places
- Support tickets in Zendesk
- Feature requests in Productboard
- Sales calls in Gong
- User interviews in Dovetail
- Social mentions everywhere
- NPS surveys in their own tool
- Churn interviews in random Google Docs

Now, they wanted AI to magically connect all of it.

And here's what we told them: AI can't fix a structural problem.

First step: pick one source, the one with the highest signal (for them, it was support tickets because every issue was already tagged and categorized).

We built an AI analysis for that one source, and it worked. From that, they got insights they could act on.

Now they're adding the others. One at a time.

If you have data everywhere, don't start by connecting everything. Start by getting value from one source first.

Client last month: "We just need you to make our chatbot smarter."

Turns out their chatbot was giving wrong answers 40% of the time.

We dug deeper, and we found out that the problem wasn't the AI, but their knowledge base:

- Outdated pricing in 12 different PDFs
- Product specs scattered across Slack threads
- Support answers that contradicted each other

No model can fix messy data. It just makes confident mistakes faster.

What we did next was to spend a couple of weeks organizing their data before touching the AI. Now the chatbot works.

If your data lives in 10 different places with no single source of truth, start there.

A founder told me they spent 6 months building the perfect AI architecture, while never talking to a single user.

They launched last month, and it turns out users needed something quite different. 

Perfect systems built for the wrong problem are just expensive mistakes. I recommend prioritizing talking to users as much as possible to ensure you build what they really need.

A16z recently led a $30M round into Lio, a startup automating enterprise procurement with AI agents.

Sourcing, contract negotiation, spend management. All of it.

I'm not surprised by the funding. Procurement is one of the most automation-ready functions in a mid-size company, and one of the most overlooked.

When I talk to operations leaders about where AI could save them time, they usually mention customer-facing workflows first. Support, sales, marketing. Procurement comes up later, almost as an afterthought.

But when we actually map where decisions are slow and errors are expensive, procurement keeps appearing: requests sitting in inboxes waiting for approval, vendor comparisons done manually in spreadsheets, contracts reviewed line by line every single time.

The pattern underneath all of it is the same: high volume, repetitive structure, clear rules for what good looks like. That's exactly where automation compounds.

The question worth asking isn't whether procurement can be automated. It's which part of your procurement workflow is costing you the most time right now, and whether the process is clean enough to hand to an agent.

If the answer to the second part is no, that's where to start.

Most enterprise clients are not ready to build AI. This would be the wrong first move, anyway.

Very often, they are excited about building AI. Budget approved, timeline set, excited to start building.

Then I look at their setup. Data scattered across 6 systems with no structure. records contradicting each other, no single source of truth.

In such cases, AI won’t be a positive addition. It will only automate the chaos.

So I get back to them with a different plan. Spend two months organizing the data first, establish structure, then build AI that actually works. 

This approach might feel slow when you're excited to start. But it’s faster than building something that breaks and spending the next 6 months trying to fix it.

If everyone you talk to says yes immediately, you're probably not getting honest advice.

AI Flow #5 is up.

This week, I'm exploring the gap between demo and production, using a client situation as an example.

They had a recommendation engine already working in production. Users were getting suggestions, and the feature was live. But it performed at maybe 70% of where it could be.

After digging into the case, I noticed some problems in how it was built. It was rigid, with hard-coded paths, no observability, and every new requirement broke previous flows.

"We built a demo that works great. Now what?"

Now you find what breaks in production. 
What you should prioritize:

- Add monitoring
- Add fallback logic
- Test with real users, not perfect test cases.

Keep in mind that the gap between demo and production is where most projects fail.

AI Flow #6 is up.

This week, I'm looking at procurement, one of the most automation-ready functions in most businesses, and the one that rarely makes it onto the AI priority list.

When we map where decisions are slow and errors are expensive across the companies I work with, procurement keeps appearing. Vendor comparisons rebuilt from scratch every time. Purchase requests waiting three days for a two-minute approval. Contracts reviewed line by line for suppliers the company has worked with for years.

Nobody designed it this way. It just accumulated.

This edition covers why this happens, what to look for in your own workflows, and where to start if the process isn't clean enough to automate yet.

Michael Mey

💼 Agile Requirements Expert | 🚀 Supercharging Productivity with AI | 🤝 Empowering Teams

Seasoned IT professional with over 20 years of extensive experience spanning diverse sectors within the industry. With roots in software development, he seamlessly integrates the approach of international
corporate environments with the agile characteristic of startups. Renowned for his structured approach, Michael possesses a keen eye for detail alongside a strategic mindset that encompasses the broader organizational landscape.

His reputation is built upon transparent communication, strong social acumen, and a pragmatic, hands-on ethos, making him a trusted advisor and leader within the digital domain. With a comprehensive skill set encompassing roles such as Project and Program Manager, Agile Coach, Scrum Master, Product
Owner, Requirements Engineer, Trainer, and Facilitator, Michael consistently delivers impactful solutions that align with business objectives.

Founding Member

Driving the New DAO Economy - twire gets your DAO off the ground and ensures long-term operations with our low-code toolkit.

Ist Programmieren jetzt nur noch eine Frage der Einstellung?

In nur drei Stunden haben wir heute eine Grenze eingerissen, die jahrelang als unüberwindbar galt. Gemeinsam mit Matthias Herbert und dem Verlag Dashöfer GmbH haben wir gezeigt: Wer ein Problem versteht, kann es heute auch lösen - ohne eine einzige Zeile Code selbst zu schreiben.

Das Experiment: Von Null zur App in 180 Minuten 
Das Besondere an diesem Seminar? Kein einziger Teilnehmer hatte einen Software-Hintergrund. Dennoch haben alle innerhalb der kurzen Zeit eine eigene, funktionierende App gebaut. Der Schlüssel war «Vibe Coding» über das Google AI Studio. Da die Hürden bei der Einrichtung und Bezahlung dort aktuell am niedrigsten sind, konnten wir uns sofort auf das Wesentliche konzentrieren: Die Idee.

Was in drei Stunden entstanden ist: 
Die Bandbreite der entwickelten Lösungen war beeindruckend und zeigt, wie individuell KI eingesetzt werden kann:
- Ein Magic Meme Creator für schnelles Content-Marketing.
- Ein automatisiertes Tool zur Erstellung von Briefing-Dokumenten für Kunden.
- Ein smarter Planer für Fussball-Trainingseinheiten.
- Ein effizienter Projekt- und Ressourcenplaner für den Business-Alltag.

Warum die Technik nebensächlich, aber das Handwerk essenziell ist 
Man könnte meinen, Software-Architektur sei nun hinfällig. Das Gegenteil ist der Fall. Vibe Coding nimmt uns das Tippen ab, aber nicht das Denken.

Was wir heute im Seminar massiv gespürt haben: Die Qualität der App steht und fällt mit der Business-Analyse und dem Requirements Engineering. Wer seine Kunden nicht versteht und Anforderungen nicht präzise definieren kann, dem hilft auch die beste KI nicht.

Im Product Management Lifecycle verschieben sich die Gewichte:
1. Das Kundenverständnis wird zur absoluten Kernkompetenz.
2. Die Fähigkeit, Geschäftsprozesse tiefgehend zu analysieren, entscheidet über den Erfolg.
3. Professionelle Software-Architekten werden wichtiger denn je, um diese schnellen Prototypen in sichere, skalierbare Enterprise-Lösungen zu überführen.

Die Zukunft des Product Managements 
Wir verlassen die Zeit, in der man Wochen auf einen Prototypen warten musste. Product Manager werden zu aktiven Schöpfern. Matthias Herbert und ich sind fasziniert davon, wie schnell aus einer vagen Vorstellung ein echtes Werkzeug werden kann, wenn man die Methoden des klassischen Requirements Engineerings mit der Geschwindigkeit von generativer KI kombiniert.
D
ie Barriere zwischen Idee und Ausführung ist weg. Jetzt kommt es darauf an, die richtigen Fragen zu stellen.

#VibeCoding #ProductManagement #RequirementsEngineering #GenerativeAI #Innovation

The LinkedIn AI bubble is a distortion of reality.

If you spend ten minutes scrolling, you’d think the entire world has already automated their lives. But this chart is the reality check we need.

The global adoption of AI isn't as deep or as sophisticated as your feed suggests.

Look at the numbers: Out of 8.1 billion people, ~84% have never interacted with AI. And for the 16% who have? The "adoption" often stops at the surface level.

In our trainings at Obvious Works, we see a recurring, dangerous pattern: People are using AI as a glorified Google. They prompt, they get an answer, and they stop thinking. They trust the output blindly without applying the critical oversight required for high-quality product development.

The tech is moving at warp speed, but our mental models are lagging.

This is why my mission is two-fold:
⦿ Professional: At Obvious Works, we help teams integrate AI across the entire product lifecycle. From requirements to implementation - without losing the "human in the loop."
⦿ Personal: As an extended board member at swissAI, I’m committed to enabling people to use these tools effectively in their private lives. Moving from "blind trust" to genuine empowerment.

AI shouldn't be a reason to outsource your brain. It should be a catalyst to think deeper.

We aren't "late" to the AI era. We are just at the stage where we need to learn how to actually use the tools we’ve been given, before we forget how to build things ourselves.

How many people in your circle are using AI as a "thinking partner" versus just a "faster search engine"?

Credit: @SoloBitcoiner (X) for this incredible visualization.

Want to move past the hype and master the practical application of AI in product management? Follow me for more insights.

#AI #ProductManagement #SwissAI #SoftwareDevelopment #DigitalTransformation

taplio

The Battle

Mihai Anton

Michael Mey

Followers

Posts

Engagements

Likes

Performance ScoreMihai Anton's LinkedIn Performance

Performance ScoreMichael Mey's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Mihai Anton

Michael Mey

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);}Mihai Anton's LinkedIn Performance

Performance ScoreMichael Mey's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreMihai Anton's LinkedIn Performance