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.

Aimsio

As a Software Developer/Enterprise Java Developer at Aimsio, I took on the challenges of planning and developing different parts of our solution including a comprehensive Budgeting and Bidding system using technologies based on Java, Vaadin framework, SQL database, message broker, and with having the MVP architecture in mind. My responsibilities included:

- Developing the Budget Planning module to seamlessly synchronize Budget items with Change Orders, resulting in an enhancement of the budgeting and dispatching services accuracy and data consistency.

- Implementing a new dynamic form editor module with the ability to have different sets of functions and views based on the user access level and role enabling users to dynamically define forms for different parts of their workflow.

- Designing and implementing a Native Spreadsheet web component for Vaadin, which enhanced data manipulation and user interaction within the system.

- Implementing a Dynamic Condition System, that controls the visibility of report actions based on the user’s session, role, and active tier, ultimately improving the system security.

- Introducing a  testing strategy roadmap for the development team, managing the migration of unit tests and integration workflows, and updating legacy JUnit 4 tests to JUnit 5 to improve the tests and code quality, and enhance the CI/CD process.

- Participating in the general system architectural design process, collaborating consistently with other team members to develop and improve the various backend and web services modules.

- Conducting instructive sessions for the development team about utilizing Ansible for automating workflows encompassing local setup updates, database migrations, and configuration tasks

I'm eager to learn and achieve more on my ongoing Journey with Aimsio!

Software Developer

University of Alberta

Worked as a Data Scientist and Research Assistant on Open Entity Resolution and Entity clustering using Deep Neural network approaches. During my collaboration, I used large language models  (LLM) and NLP techniques developing software to study their practicality on Entity Matching under the supervision of Dr. Davood Rafiei.

Teaching Assistant in File and Database Management (CMPUT 291) course under the Supervision of Dr. Rafiei. In this course, we helped students with basic concepts in computer data organization and information processing, entity-relationship model; relational model; SQL and other relational query languages; and storage architecture. They learned how to use Python, SQLite, and MongoDB.

Teaching Assistant in Information Retrieval Course under the instruction of Dr. Denilson Barbosa.

Applied Data Scientist | Research Assistant

FANAP

At Fanap Co., I was a part of the SAKKU Container Orchestration Platform and POD Space Cloud Storage teams. These experiences allowed me to work closely with my colleagues to address complex technical challenges including:

SAKKU Container Orchestration Platform
- Worked as a software engineer leading a diverse engineering team with 9 software developers and 2 DevOps engineers.
- Designed the project's architecture, implementing microservices software using Spring Framework and MVC architecture.
- Hands-on involvement in backend development, fixing bugs, adding new features, and improving performance.
- Planned the project's technical roadmap, guiding its evolution in line with our strategic objectives.
- Leveraged cutting-edge technologies including but not limited to Apache Marathon, Mesos, Docker Swarm, etc in the development process, resulting in a remarkable 70% reduction in resources.
Open-sourced many developed tools, services, and Docker images to support the open-source community. 

POD Space Cloud Storage 
- In the POD Space Cloud Storage project, I had the opportunity to contribute as part of a great team, solve complex problems, and collaborate to address various technical challenges. As a software engineer, I played an integral role in enhancing the cloud storage module, refactoring it using Spring Boot, and implementing an efficient file-sharing schema that resulted in a 40% improvement in performance.
- Created a scalable service to generate thumbnails for the uploaded files.
- Our team successfully migrated the project from Java EE to the Spring framework, standardizing the REST APIs and optimizing the response time performance.
- My involvement extended to designing and developing a robust file-sharing system, implementing APIs for seamless file-sharing, and defining access levels for files and folders.

Software Engineer

WeClick.ir

Tapleader SDK was an in-app real-time crash reporting platform that helps mobile applications prioritize and troubleshoot their most prevalent crashes based on the impact on real users. I implemented the Android library with Java and Android SDK to collect clients' activity data in different environments.

SignalGo is a framework to distribute a  computation among a group of mobile clients. My task was to design the client-server architecture of the framework and implement the Java client using Java 8.

I am a dedicated and experienced software engineer with a strong focus on Java web development frameworks, backend development, and containerization technologies. Over the course of +8 years, I have honed my skills in creating robust and scalable solutions for complex challenges.

taplio

The Battle

Mihai Anton

Matt Akbarian Rastaghi

Followers

Posts

Engagements

Likes

Performance ScoreMihai Anton's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Mihai Anton

Matt Akbarian Rastaghi

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