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.

AD Product Experience Manager

Product Experience Intern for Autonomous Driving

Innovation Labs Romania

Innovation Labs - the biggest university pre-accelerator for tech startups in Romania. In other words: we help define ideas, build strong team connections, educate in business and shape mindsets - right from the get-go!

- Ensured constant communication with all our cohorts, adapting the content, tone of voice and communication channels, as well as increased our newsletter open rate by 20% since 2022.
- Crafted social media copies and graphical designs to promote and keep the online presence.
- Organized and supported our onsite events with logistics, social media strategy and event flows.

Head of Community Engagement

- Digitalized the hackathon and program experience of the biggest startup pre-accelerator program in Romania by migrating it to Discord, which included structure and setup, including bot configurations and custom commands, role and permission management. 
- Designed the registration and hackathon online processes.
- Conducted workshops with the entire national team to adjust from onsite to online events on Discord.

- Designed visuals and wrote copies for social media, managing several accounts together with the team.
- Organized and communicated relevant details to the current program cohort.
- Supported with online and onsite logistics.

Logistics and Communications Specialist

Invibes Advertising

Worked on a new product together with the Product Owner in an Agile team, using Jira and Confluence for story management and documentation. As the only developer, I was responsible for designing the project architecture, choice of frameworks, and code writing. Involved in the UI/UX design.

Responsible for adapting one of the company's platforms to a new mobile app,
including UI/UX design and application architecture.

iOS Developer

Worked in an Agile team consisting of iOS, Android, backend developers, and testers on a mobile banking application (100K+ downloads) using Swift with UIKit and following the MVVM design pattern.
My responsibilities included new features development, bug fixing, technical improvements, Unit testing.
Contributed to a better user experience by suggesting small changes that made a difference in the flows.
Was the main iOS developer responsible for AMS, building, deploying and helping with testing the app.

Other responsibilities inside IBM Romania:
- Scrum master for an iOS internship team: conducted Agile ceremonies and refined the backlog
- Facilitated design thinking workshops for internal events
- Got involved in several internal initiatives, including facilitating ideation workshops and developing internal communication strategies, some as part of my Awareness Stream Leader position inside the "Lead your Leader" community

- Attended workshops on mobile app development, Swift and Agile
- Worked in a team of four interns to develop an app from scratch using UIKit and MVC design pattern, simulating the work in a real-life project following the Agile mindset

iOS Developer Intern

BEST Bucharest

Together with the other four members of the organization's board, I created the yearly board strategy based on the needs and wishes of our local BEST group and held an overview of all the ongoing activities. This also meant making sure that every member feels appreciated and encouraged to try their own approaches in the area of complementary education and self-development.

As the PR responsible, I created and coordinated the image strategy throughout the year, took part actively in the development of all PR materials (online & offline), directed the information flow on social media and the promotion activities of the organization. My duties were also to maintain contact and collaboration with our media partners and fellow student NGOs.

Throughout the year I had an overview of the departments PR, Design and IT, maintaining constant communication and idea exchange with the three department managers. 
Alongside my board, I supervised 3 successful events, 7 core teams, an 8-member management team, and adapted our strategy and perspective mid mandate due to the pandemic, maintaining our volunteers engaged.

It is one of the most treasured experiences that taught me how to better work in a team, prioritize the tasks and really value the people around me.

Vice President for Public Relations

Capgemini

Got acquainted with the Cucumber Framework and developed a POC to perform specific searches on Google and Youtube.
Trained for the ISTQB certification.

Automation Tester Intern

Hi! I’m a computer science graduate who is passionate about exploring the intricacies of human life and behaviour. Equipped with an engineering background, I have now embarked on a journey in the product experience sphere, creating user-centric products. 

As Head of Community Engagement, I'm also part of the 12-member core team at Innovation Labs Romania, the tech start-up pre-acceleration program!

I'm grateful for my student volunteer years at BEST Bucharest, where I honed skills like teamwork, communication, leadership, ethics, and critical thinking. Outside of work, I enjoy painting, playing the piano, reading, hikes, and exploring the cultures of the world as much as I can.

I've recently started a personal collection of formidable words in any language - I'd love to hear your favourite word and its meaning!

taplio

The Battle

Mihai Anton

Iulia Moroti

Followers

Posts

Engagements

Likes

Performance ScoreMihai Anton's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Mihai Anton

Iulia Moroti

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

Ready to grow your LinkedIn presence?

Performance ScoreMihai Anton's LinkedIn Performance