AI Flow

• Working with small and medium companies, building deep tech AI products from 0, extending existing projects, and taking 100% ownership of the technical part, so you can focus on your business. 

• Having a deep understanding of AI and Machine Learning, I also do consulting for high-growth companies. Working with Python, Typescript and C++.

• Clients include BP, ExoMatter.ai, colayer.io, deinestudienfinanzierung.de and Interflexion.

Founder, Lead ML Engineer & AI Consultant

Metaphysic.ai

• In charge of the ML team. My main goal is to productionize research projects and lead the team towards longer-term goals.

Lead Machine Learning Engineer

• Working on generative AI for hyperrealistic video content creation. Building machine learning models, pipelines, and infrastructure for training and inference.

• Reduced the existing ML pipeline time and cost by 45%.

Senior Machine Learning Engineer

Uplift.ltd

• Worked on AI, Cloud,  Data Engineering, and Full Stack projects. Handling the whole product lifecycle, including planning, implementation, and deployment.

• Clients include interflexion.com, collectoralerts.com, and perfectlypitched.co. I worked with them to build features in Node.js, Django, Next.js, and React, and scaled their infrastructure on Google Cloud. Built apps from scratch with Next.js - GraphQL}- PostgreSQL, deployed on Digital Ocean.

Senior Software Engineer

Google

• Did research on feature selection and built a tool that makes Machine Learning iterations up to 70% faster. Clients included Ads and Cloud.

AI Engineer | Python, PyTorch & Tensorflow

• Improved the user experience in Search and Maps, by building data analysis tools for Geo Location. Used C++, MapReduce, and SQL to build a distributed data pipeline for analyzing and fixing locations with issues.

Software Engineer

Bloomberg LP

• Built a sharing feature in the internal collaboration tool with 200K daily users. Worked Full-Stack with JavaScript and TypeScript, leading the project from start to delivery.

• While in the News Feeds department I moved existing services to Docker, making development flows smoother, and impacting more than 300 engineers. I collaborated with stakeholders, identified their needs, and implemented the migration.

• Built a cache infrastructure for mapping users to communities that currently serves 200 requests/second, significantly improving the loading time client-side. Used C++ and Python for development, Redis for building the cache layer and RabbitMQ for listening to database updates. Collaborated with the Sharing Framework to implement and integrate the cache infrastructure with their existing system.

Software Engineer Intern

METRO AG

Bosch

• Worked on anomaly detection in videos containing thermal events. Starting from unlabeled videos of running circuit boards, built a classification model which decided, for each frame, whether or not it had an anomaly (contained fire or smoke). Used Keras for building the Convolutional Neural Network.

Data Science Intern

Bitdefender

• Worked on a data structure library in C (hash tables, binary search
trees, heaps, linked lists, and arrays).

Innovation Labs Romania

• Built an Android application for electric scooter sharing. My team was among the best ones which granted us a mentorship program to further develop our product.

Innovation Labs Hackathon 2019

Senior Machine Learning & Full Stack Engineer • 9+ years in AI & ML • I built AI at Google, Bloomberg, and early-stage startups • Founder @AI Flow | We build end-to-end data products • AI Consultant • Open to B2B, remote projects.

 • I've been in the software industry for more than 8 years, splitting between Software Engineering and AI. After spending some years at Google and Bloomberg solving problems at scale, and building products for startups, I decided to build AI Flow, an AI focused Software Agency. We work with small and medium companies, building deep tech AI products from 0, extending existing projects, and taking 100% ownership of the technical part, so you can focus on your business. Having a deep understanding in AI and Machine Learning, I also do consulting for high growth companies. Let's get in touch!

 • In my day-to-day work, I am:
☑️ Designing and prototyping large-scale ML &  data systems.
☑️ Writing clean Python, C++, and Typescript.
☑️ Leading and managing high-performing teams.
☑️ Communicating clearly and often.
☑️ Building and training models with Pytorch.
☑️ Developing full stack with NextJS, Node, React, and Django.
☑️ Deploying on any cloud with Docker and Kubernetes.
☑️ Focusing on quality, simplicity, and a great UX.
☑️ ... and many more.

 • Only interested in remote contract opportunities •  Book a meeting on https://antonmih.ai/meet

 • Website: antonmih.ai • Phone: ‭+40 721 658 931 •  Mail: mihai@aiflow.ltd

Mihai Anton

Andrej Karpathy got early access to Gemini 3, and the model refused to believe him when he said the year was 2025.

Quick context if you don't know him: Andrej Karpathy is a founding member of OpenAI, ran AI at Tesla, and now builds AI-powered education tools. He knows how these systems work better than most.

So when he told Gemini 3 it was November 2025, and the model accused him of lying, things got interesting.

He showed it in news articles, images, and Google search results. The model dug in harder, claiming he was uploading AI-generated fakes and even pointing out the "dead giveaways" that proved it.

Turns out the model's training data ended in 2024, and Karpathy forgot to enable the Google Search tool. The model was essentially disconnected from the world, defending what it knew with total confidence.

Once he enabled the search tool, Gemini 3 saw 2025 for itself and immediately apologized. It even marveled at current events: "Nvidia is worth $4.54 trillion? This is wild."

The lesson here: models operate purely on patterns from their training data. They don't have a concept of "current reality" versus "outdated information." 

Without retrieval systems or live context, a model will defend that we're living in 2024 with the same confidence it would defend that we're living in 2025. Both are just patterns to it.

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

Mic check done. Seats are getting filled. Looking forward to see how the surprise I prepared for the audience lands, but I’ll share more about it after. 

TEDxBrașov, excited for this!

(8) Feed | LinkedIn

JSESSIONID

li_at

Talks about #ai, #startups, #remotework, #freelancing, and #softwareengineering

Off to Dubai for the next couple of days. Catch me at the World Tech Summit! 

If you’re around and into AI, engineering, or just good coffee + tech talk, let’s meet. I’d love to connect with fellow builders, founders, or anyone curious about GenAI beyond the hype.

So, let’s grab a coffee and talk about AI. DMs are open.

Travel can be a powerful add-on to your work, as long as you have the right systems built for it.

Switching environments (without dropping structure) helps me reset my thinking, notice weak spots, and come back to problems with a sharper eye.

This week, I’m balancing work with attending the World Tech Summit. That means:

- Deep work blocks stay protected (calendar is the source of truth).
- One problem per block, written in one line before I start.
- Async routines keep moving in the background.

New scenery, same structure, sharper focus.

No alternative text description for this image

Three months in Cluj to turn “I’m testing an idea” into “people are using my product.”

I wish I had this program when I was 20.

If you’re under 25 and already building in your spare time, skip the CV part and come test your ideas properly. You’ll get time (a monthly allowance), room to experiment (a real ops budget), and a real prize on the line (an investment from @ Lost Astronaut VC).

It’s supported by @Cluj Startups, so you’re not just in a program, but in a community. Their office, Founders Hive, is your home base during the program.

If you join, tag me. I want to see what you build!
Apply here: thebasejump.com/arena

Some of the best decisions I’ve made came right after I paused.

Not when I pushed harder, or when I tried to force a breakthrough. But when I stepped away for a few minutes, cleared the noise, and let the problem breathe.

In tech, it’s easy to mistake urgency for progress, as it might seem like you’re always one sprint behind, one release away, one Slack message from making the next move. But in architecture, product design, or hiring, rushed decisions have a long half-life.

It’s rarely the loudest thought that’s the right one. Most often, it’s the one that still makes sense after a walk.

I see a lot of AI startups treat architecture like something you "fix later." The mentality is: ship the MVP, get traction, then rebuild properly.

Here's what actually happens:

Your MVP works for 50 users. You get traction. Now you need to scale to 500, then 5,000. But your system was built for a demo, not for growth.

So you're rewriting core infrastructure while trying to keep the lights on, which means you're patching instead of scaling. And every new feature takes twice as long because the foundation is shaky.

The time you "saved" by skipping architecture? You're paying it back with interest.

I'm not saying over-engineer, but I encourage you to think about these questions first:

- How will this handle 10x the load?
- Where will the bottlenecks show up?
- Can we monitor what matters?
- Is this decision reversible, or are we locked in?

The best teams I've worked with move fast AND build clean. They know which corners are safe to cut and which ones will cost them later.

Who is asking to “add GenAI,” and how you can shape the conversation:

A) Tech lead

Ask: What bottleneck do you want to eliminate next month?

If they can show a metric, design a scoped assist. If not, run a one-week discovery.

B) Founder

Ask: What outcome are we missing, and what would prove it?

If they can pick one metric, design a pilot. If not, map the value chain and choose the highest-friction job.

C) The “We need AI” stakeholder

Ask: Which workflow, and what value per run?

If they cannot answer, pause the build and run a process audit first.

Always remember that tool choices come last, while clarity (on job, outcome, and metric) comes first.

OpenAI is aiming for an “intern-level research assistant” by 2026 and a “legitimate AI researcher” by 2028. It’s a big milestone, sure, but I believe the most interesting part is beyond the headline. 

Let’s dig deeper:

- “Intern by 2026” reads to me like: a model that can read papers, compare findings, and suggest next steps. This implies it can do useful analyst work, still supervised.

- “Researcher by 2028” suggests autonomous experiment loops with human oversight. This could be similar to a colleague you manage, not just a tool you poke. If they land anywhere near this, the leverage will shift from prompt cleverness to problem framing, data plumbing, and evaluation discipline.

What I recommend to anyone serious about agents to do:

- Treat every agent step as a hypothesis. Add validators, timeouts, and rollbacks so progress is earned, not assumed.

- Build evals early. Keep a living regression suite with clear failure buckets (fabrication, tool misuse, spec drift).

- Get provenance in order. Lineage, permissions, and auditability beat “move fast” when an autonomous loop makes a wrong turn.


This way, you’re already designing for trust. When capabilities unlock, your system is ready to absorb them without chaos. 

Guardrails, evals, and data lineage aren’t nice-to-haves anymore. They’re how you convert tomorrow’s “AI researcher” into reliable throughput today.

There’s real gold hiding in failed sessions. Your job is to replay, map, and fix the first mile.

graphical user interface, text, application, chat or text message

Too many AI projects still start with “Which model should we use?”, instead of:

- What’s the real problem we’re solving?
- What decision needs to be improved?
- Where does it fit in the system?

This results in a parade of flashy demos that never ship, because there was no strategy to begin with.

LLMs are amazing tools, but they’re not strategy. 
They don’t know your business. You do. Start there.

We spend a lot of time discussing model accuracy and deployment speed, which are crucial. But one of the most underrated technical decisions, and a critical safeguard for sustainable AI, is designing for the "Human-in-the-Loop." 

This isn't a sign of AI weakness; it's a strategic strength.

Humans bring judgment, contextual understanding, and ethical oversight that algorithms simply can't replicate, especially in ambiguous or high-stakes scenarios. 

Whether it's validating model predictions, handling edge cases, or providing feedback for continuous learning, integrating human expertise at key decision points enhances accuracy, mitigates bias, and builds trust. 

For any Mid-sized company with Data looking to formalize AI, or an AI Startup building mission-critical products, prioritizing this interaction ensures your AI remains reliable and accountable.

Reliable agents come from clear lifecycles, tight scopes, confidence gates, and built-in observability.

“Agents” look great in a demo. Then they meet real work: partial context, fuzzy ownership, and failures that don’t announce themselves. What helps is treating agents like operators in a system, not like magic.

What actually breaks

- Missing state. Acting without the latest facts or forgetting what matters.
- Ambiguous ownership. Nobody is sure who approves, who fixes, who decides next.
- Silent errors. Things go wrong quietly and, the customers find out first.

How to make them dependable

- Name the lifecycle. Write the states down (queued → in_progress → needs_human → done → audited) and stick to them.
- One job per agent. Clear intent, clear tools, clear output.
- Confidence gates. When certainty dips, pause and escalate.
- Human checkpoints for high-risk moves. Review before actions that are costly or hard to reverse.
- Build in observability. Log inputs, tools used, decisions, and the evidence behind them.

A quick readiness test I use:

- What is this agent allowed to do, when, and with which proof?

If that isn’t crisp, shipping will just move the ambiguity into production.

Agents don’t fail at “intelligence”, they fail at handoffs. 

Who are the usual suspects?

- Missing context → the next step is a guess.
- Ambiguous ownership → nobody rescues stuck work.
- Silent failures → confidence drops and no one knows.

Here’s how you can design around it: one job per agent → explicit states → confidence thresholds → human gates on high-impact steps → clear recovery.

Then, make sure to track and learn from the insights. Start by looking at four numbers: handoff success rate, time to first correct handoff, intervention rate, and silent failures (target: zero). If those improve, you’re on the right track. 

Your model might not be bigger, but it surely has cleaner edges.

Another release won’t help if no one owns the outcome. Pick one owner, pick one number.

Make ownership visible in how the work runs. 

Here’s the checklist I use to wire it in:

- One owner per workflow. One name, one inbox.
- One number that fits the goal.
- Ship to usage. “Launched ✅” doesn’t count until people rely on it.
- On-call = accountability. If you build it, you take the call when it breaks.
- Weekly user time. 30 minutes with real users > 30 comments in a doc.
- Rollback planned upfront. Know how to back out before you push.
- Monthly teardown. What worked, what hurt, what we stop doing.

When ownership is clear, priorities stop drifting and feedback loops stay alive.

What actually moved the needle this year? What quietly got in the way?

I put together 4 paste-only prompts you can run in any AI chat. All you need = your notes as inputs. The answers will be honest, and after running all 4 of them, you should also have a plan for your next 12 weeks. 

Treat December as the cleanup month: spot inconsistencies, clear the friction. This way, you can start 2026 with the insights you need.

Save this carousel for a quiet hour. 

I’ll come back with a version for businesses soon.

Don’t scale the unknowns

If one user run feels fuzzy, a thousand will amplify the mess. That’s the right moment to slow down and trace the path: inputs → decision rules → next action → owner → evidence. 

Make the edges crisp at 1×, then turn the dial.

“We want to use our internal data for AI.”

Cool. But... why?

One mistake I see over and over: teams don’t define success metrics before they start. It’s treated like you can throw your Google Drive into a magic AI bin and expect results.

AI isn’t magic. It’s a tool. And tools need a purpose. So, are you trying to:

- save time on a workflow?
- cut operational costs?
- lower response times for your support team?
- improve the quality of content or insights?

Or is your product itself AI‑driven, and leveraging internal data is the only way to serve your clients?

Without a clear “why,” you’re just experimenting in the dark, and it shows when you try to measure impact later.

Before you start building, get specific about what “better” means for you. Your data (and your team) will thank you.

What success metrics do you look at before kicking off an AI project?

Are you building for the next two years or the next two weeks?

Most teams say long-term, but their roadmaps disagree. If a routine model upgrade, a traffic spike, or a vendor price change knocks things off course, you’re just renting stability. 

I always encourage teams to make three decisions early on:

- First, what must be resilient: identity, data contracts, evaluation harness, and billing. Define SLOs, error budgets, RTO, and RPO. Be explicit about how each part fails and how it recovers.

- Second, what should stay swappable: model provider, embeddings, vector store, feature flags, analytics SDKs. Create clean seams, hide them behind interfaces, and keep a tested fallback that preserves core function even at lower quality.

- Third, what you will outsource: queues, authentication, and observability. They are essential, but they rarely set you apart. Spend engineering time on the paths that create user-visible value.

Which of these three decisions have you postponed? What would break first as a result?

Personalization isn't a model problem. It’s a data problem.

The model can only work with what you feed it. And most teams feed it scraps:

- Public datasets
- Generic CRM data
- Surface-level attributes

And then they wonder why their results look like everyone else’s.

The teams winning the hyper-personalization are:

- Investing in collecting meaningful signals
- Labeling their data for depth, not just volume
- Building internal pipelines, not just prompt wrappers

You don’t need a bigger model. You need better input.

taplio