Epoch8

Epoch8.co – is an agency for data analytics and machine learning projects.

Co-Founder, Head of Analytics

Эвотор

Business Development Manager

Mail.Ru Group

Product Manager

Readymag

Product Analyst

Enter.ru

Enter is one of the largest Russian multichannel retailers, having 80+ offline shops, e-commerce site, sales terminals and mobile applications.

Responsibilities:
Product Management:
— Product and UX development planning and roadmapping;
— Detailed new features description: product concept and UX wireframing;
— KPIs definition;
— Key projects supervision;
— Analytics (users behavior and business metrics).

Team Management:
— Managing the product team: 3 product managers, responsible for web site, mobile applications and CMS. My team coordinates work with ~20 web developers and content editors.

Head of Product Management

News360

News360 is a personalized news reader with more than 3 mln users across multiple platforms (iPhone, iPad, Andtoid Phone, Android Tablet, Windows 8, Web). News360 platform gathers news articles and blog posts around the web and, by learning what users read the most, brings them content that they’ll find interesting and important. 

News360 app was featured multiple times in both Apple appstore and Android market.

As a PM of News360 team (located in Moscow and San Francisco) I managed development of iPhone, iPad, Android Smartphone, Android Tablet, Windows 8 and Web versions of News360 app, including:
— working with business owners for business requirements gathering, analysis and specification;
— project task decomposition, scheduling and tracking;
— QA-management and release management;
— post-launch analytics and improvement: core KPI metrics definition, feature usage etc (using Google Analytics, Flurry Analytics, internal tools).

Achievements:
— 50+ releases of News360 apps across all platforms, including two super-major total redesign releases;
— introduced regular process of product research, development, testing and releases.

Project Manager

Undev.ru

I worked for technologically challenging TV and video broadcasting projects.

Completed projects:
— P2P video broadcasting platform (project manager);
— Information system for the St. Petersburg Economic Forum, 2011 (analyst).

1. P2P video broadcasting platform

Internal platform supporting video broadcasts organized by Digital October.

Technologies:
— TorrentCDN: content delivery system upon Bittorrent protocol (C++; Erlang);
— Server-side application for video capture and slicing (C++);
— Desktop applications for P2P video streaming (Windows, Mac, Linux).

2. Information system for the St. Petersburg Economic Forum, 2011
The information system allows operator to control 200 Forum video panels: schedule video broadcasts from conference rooms, display media content, real-time Twitter feeds etc.

Objectives:
— business requirements gathering, analysis and specification.

Project manager / Analyst

Kaspersky Lab

I was responsible for increasing online sales of Kaspersky products (Russia and EEMEA).

Objectives:
— Omniture SiteCatalyst implementation, working with Omniture to determine and improve typical user paths and conversion funnels; AB-testing of multiple page variations to improve conversion.
— Order process redesign, including both web site UI and business logic, resulting about 23% conversion lift.

Internet Solutions Manager / Web Analyst

Freelance

Web projects management from scratch to post launch analysis for big Russian media and FMCG brands (Il Patio, 1-2-3cafe, Media Arts communications group and some others).

Freelance Project Manager / Analyst

HeadHunter Group

UX designer / Product Manager

I'm a co-founder of Epoch8.co and Agima.ai, specialist agencies for machine learning and data analytics projects.

Why you might consider talking with us:
– Data analytics: if you have a startup and need a proper technical setup of all the analytical infrastructure. We help to set up ETL/ELT pipelines, gather data from all data sources, clean it up, describe it and prepare for reporting and self-service usage. Our focus is to do preparatory work to make it easier for the product teams to create reports quickly.
 
– Machine Learning: if you need Computer Vision functionality in your mobile app, for example recognizing details, photos, videos, people, signs and symbols, gestures, etc. We run ML models in real-time on the mobile device, so the user gets an experience of lightning-fast object detection.

We’ve been building an AI assistant for a client in a highly regulated industry. For almost a year.

And in these kinds of industries, there’s a funny thing:
- You’ve got the law written in documents;
- And then you’ve got how experts interpret the law, which lives in people’s heads.

And those interpretations are hidden.
After each test, you get lo-o-o-ong discussions:
- The expert says, "No, the AI is wrong here"
- You go, "But why? The law literally says so"
- And the expert goes, "Yeah, but in practice…"

Let me give an example.
Say you manufacture some product. You need to assign the correct legal codes to it.
There’s no "search for a product name and get the code" solution.
It’s more like:
"If it’s a household item, made of plastic, but only if it’s not for industrial use – use code X.
And if it’s made of wood and is for industrial use – then use code Y".
You get the idea.

So.
We spent a year testing and debugging.
→ We turned legal rules into rows in a table.
→ Broke out conditions into separate fields.
→ Taught the AI to follow the logic.
→ And finally started getting decent results.

We had a team debrief. Everyone agreed:
"Yeah, yeah, just dumping documents into AI doesn’t work. You have to prep the data, clean it, explain the logic, etc etc"

End of meeting.
Then the CEO says: "Great. Send me your wishlist for AI features for the next year. Let’s plan our next steps"

Next day, I open this wishlist.
And here’s what I read, after a whole year of real experience:

- TWO people wrote: "I want an AI where I can just upload files and it answers everything"
- And one person requested: "astrology-based communication strategies with clients. Like, it knows they’re a Virgo and adjusts precision level"

And honestly, that second one might actually be more realistic 😅 

[Photo creds: me, with my hair having a nervous breakdown after untangling legal data so the AI could chew on it]

Olga Tatarinova

We help founders and internal teams build computer vision, data, and AI products ❤️ eCommerce & retail, warehousing & logistics, LegalTech, HealthTech, startups

We finally sat down and gave our main Epoch8 deck a proper refresh – added new projects and clarified what we *actually* do.

It now covers:
- computer vision (incl. real-time mobile pipelines),
- precise AI assistants,
- AI/ML Proof-of-Concept lab for fast experiments,
- tracking & analytics clean-up,
- typical pricing.

Here it is if you’re curious or want to share with someone who’s scouting for a solid tech partner.

Also: we’ve got CV and AI engineers available starting January. If you’re kicking off something ambitious (or just need quick help) – we’re around.

Epoch8 services & case studies

I knew we were turning into a brand when I heard a client describe us like this:

"Ah, Epoch8, I know them. Family shop of three Tatarinovs: two normal humans, one ex-Googler, and three cats"

Co-founder @ Epoch8 AI/ML agency · We help founders and internal entrepreneurs launch real products built on computer vision, data, and AI · ❤️ eCommerce & retail, warehousing & logistics, legal, health, startups

Don’t let AI "just do it." Make it plan first.
We do a lot of client work, and effective AI design patterns always catch my eye.

One pattern I love is Structured Chain-of-Thought, and it works like this:
→ Task by human →
→ Plan by AI (as a separate doc) →
→ Review by human →
→ Execution by AI →
→ Final check

[Why it works]
If you throw a complex task at an AI and let it "just do it", it can invest a lot of time and tokens in an ineffective solution (like a junior employee left alone).
But if you ask it to plan first, it gives you a clear breakdown of what it’s about to do.
Now you can review, fix, and approve that plan before any real work begins.
This single step saves hours.
At that point, the AI becomes (more or less) reliable – AI is actually good at following the plans.

We saw this firsthand with Antigravity, Google’s new code assistant.
It bakes this pattern into the workflow:
- AI generates a structured Plan document
- You review and edit it
- Once approved, the AI turns it into its internal to-do
- Then it builds

Andrey Tatarinov said: "Honestly, I wish our juniors worked this way".

It’s the same in personal productivity by the way.
Procrastination often comes from starting a task without knowing what "done" looks like.
If you’re stuck, don’t force the doing – stop and make a full plan first.
Once the plan is on paper, execution becomes easy. You’re just following steps.

[Takeaway] 
Structured Chain-of-Thought: Plan → Review the artefact → Execute
Ask AI to write the plan. Fix it. Approve it. Then let it build.

We all tried it: dump a pile of docs into an AI assistant, ask a question, and get a confident, elegant – and completely wrong – answer.

Here’s the math behind why that happens.
And what to build instead.

Docs In ≠ Answers Out

I spend most of my time thinking about learning – it’s literally how I make a living.

Not only machine learning, but how people learn and unlearn behaviors.

And the more I look at it, the more I notice the same pattern:
When something becomes *too important*, both people and models start performing worse.

What’s interesting is how similar this distortion looks in human brains and in LLMs.

Let me show you what I mean.

The moment something becomes "too important"

In 2025 the limits of artificial intelligence are becoming clearer.

I talk a lot with clients about AI and LLMs, and someone asked a good question:
Is there another big jump coming? Or are we stuck on a long flat phase?

Like with space flights – in the 1960s people thought:
"In 20 years we’ll live on Mars".
But here we are, it's 2025, and still no Mars cities.
So maybe AI is the same?

Here's what I think.
Right now, LLMs have a core problem – they don’t have a model of the world.
At their core, LLMs are very strong next-word predictors, like a very good autocomplete.
They read tons of text, so they know a lot of facts.

But they don’t really track:
- what the world looks like right now
- how actions change the world
- how time passes
- what causes what

They don’t think like:
"I did X → the world changed → now Y is possible"

Why does everyone suddenly talk about "world models"?
Because without them, AI is bad at:
- doing things step by step
- understanding time and cause
- planning ahead
- controlling robots or machines

If AI had a world model, it could:
- play scenarios forward
- test plans before acting
- work better in places where the world changes all the time (like streets and roads)

That’s why big companies shifted focus in 2025.
For example:
- DeepMind announced Genie 3 – models that simulate physical spaces like warehouses
- Meta anounced V-JEPA 2 trained on videos, to learn how the physical world works
- NVIDIA anounced NVIDIA Cosmos – a platform that allows creating digital copies of real environments.

But do we really need a world model?
People argue about this.

Camp one says:
Without a world model, there is no real understanding. We will hit a ceiling.

Camp two says:
LLMs already have a weak world model hidden inside their weights.
They know water is wet. Things fall down. Fire burns.
Maybe big models plus tools are already enough for most real tasks.

I think the truth is probably in the middle.

My bet:
- The next important breakthroughs won’t be about bigger chat models.
They’ll be about memory, planning, learning and models of how the world changes.
That’s where things get interesting.

***
Image credits: screenshots from Google DeepMind, Genie 3 (source: DeepMind blog).

These are AI-generated interactive environments built by a world model.
The model:
- sees the world from a first-person view
- understands actions (move, fly, paint, avoid obstacles)
- updates the scene as actions happen
- keeps basic physics and continuity over time

This is what people mean when they say "AI needs a model of the world".

I’ve worked with 50+ founders, product teams, and engineers this year, exploring AI use-cases and the actual value it brings to business. 

Here’s how they make progress with AI instead of getting stuck in endless PoCs:

They start with one clear use-case backed by clear data.

Step 1. Pick one use-case that saves time or money today.
Step 2. Write down the main nouns: the important "things" in your domain (like "counterparts", "projects", "bills").
Step 3. Clean a small slice of data for these things.
Step 4. Build the smallest setup that answers one real question.

That’s how you ship your first AI helper without a big data team or heavy infra.

If you want faster output, fix one workflow and remove extra manual steps.
If you want fewer errors, give your models clean data to work with, not a pile of messy docs.
If you want steady progress, block one stable slot each week to close the next small gap.

Give it 4-6 weeks, and the shift becomes clear.

I’ve seen this many times: teams think they need one giant assistant that can answer every question about every document and every process – finance, HR, projects, contracts, anything. 
But once they clean one small corner, the rest stops feeling impossible.

And the classic xkcd chart fits here perfectly 👇

Co-founder @ Epoch8 · We help founders and internal entrepreneurs launch real products built on computer vision, data, and AI · ❤️ e-com, logistics, legal, health, startups · Sharing lessons from 200+ projects

Wanna burn cash? Skip this post. 

The best part of my job is helping founders and internal leads shape their AI/ML product ideas before building.

One thing always gets overlooked:
- Everyone plans how much it’ll cost to *build*.
- Almost no one asks how much it’ll cost to *run*.

This week I spoke with a team planning an AI service generating images and voiceovers for books.

I brought them a breakdown of infra and experimentation costs… and they got really quiet.
LLM tokens, image generation, audio processing, cloud bills – it stacks up fast.

But there is a workaround: most teams don’t realize how many generous startup programs are out there, covering infra, LLM tokens, image generation and TTS models.

If I were in their shoes, here’s what I’d do today:

✅ Infra
I would start with getting the Big Three startup offers for cloud infra: Google Cloud / AWS / Azure.
If you're ok with Azure Cloud, Azure for Startups grants $5k in credits, with no VC backing required and almost no questions asked.

✅ LLMs
I would use OpenRouter to access free LLM models for experiments – they have Qwen, GPT‑OSS, Mistral, Llama and some others available for free with rate-limits.

And if I had fast and dumb tasks like text summarization (no deep reasoning), I might use Groq’s free developer tier for fast generation – they are famous for their speed (generation at ~2000 tokens per second).

Also, I would consider applying to the Groq for Startups program – they grant $10k in credits if accepted, also no VC backing required.

✅ Image generation
For images, I would use Pollinations AI – a free, open platform for text and image generation. They offer free access with watermarks and are fairly prompt-sensitive, but it's enough for early experiments. 

✅ Voice (TTS/STT)
There are also three top generous programs for startups in this space:

- ElevenLabs – free access for one year for startups, if accepted
- AssemblyAI – 200,000 hours of speech-to-text and voice AI
- Deepgram – up to $100k in credits

None of these programs require VC backing, and they’re absolutely worth applying to if you're building a TTS/STT app. 

***

These won’t solve everything, but they’ll buy you time to learn, test, and focus on what matters: building something users actually want.

***

I’m curating a full list of AI/infra perks for early teams – planning a November update.

Want the full list? Comment free stack and I’ll send it to you.

There’s a common problem in warehouses and sorting centers.

You have boxes, parcels and pallets moving across the conveyor lines.
And you have cameras everywhere.  

From time to time someone asks:
"What happened to this specific item?
It was moving along the line, everything looked fine - and then it wasn’t found where it should have been".

When this happens in real life someone usually sits down and scrolls through the video footage by hand, trying to find the last moment where the object was still visible.

Recently Meta released *SAM 3 (Segment Anything with Concepts)* computer vision model, and it’s a great example of what's already possible if you look around.

At a basic level, it can:  
- recognize all objects from a prompt (for example, you type "cardboard box" and it gets it without manual image labelling),  
- separate identical-looking objects from each other,  
- and track each individual object over time, even when it moves, rotates, or gets partially covered by others.

If you watch the demo video with boxes, it becomes very intuitive.  
You can now follow *this exact box* across the video stream and see where its story ends.

This opens the door to much better camera-based analytics. Of course, when you need tracking across different cameras, this will still require extra logic or even operator intervention, but as a building block, SAM 3 is already very strong.

In practice, most business processes  work best when people and systems join forces.  
An operator points to a specific parcel on a video, and the system helps answer the hard part: where it was last seen. 
That alone can cut hours of manual video review down to minutes.

---
We’ve built computer vision logic behind tools like ACI - AI-powered Control and Inspection and keep exploring practical CV use cases for warehouses and sorting operations.

If you’re dealing with similar challenges or know a team that’s thinking about warehouse automation with computer vision, I’d really appreciate a conversation or a warm intro.

I’ve been thinking a lot about team dynamics lately.
And how different our natural styles can be.

Andrey states his method like this:
"I throw people at the problem. And I come back only when the fire’s either out or out of control"
Said with a straight face. No irony.
 
Wrote about this phenomenon here: https://lnkd.in/dgT96k6s

As for me – I can’t.
I keep circling this one idea I read somewhere:
The strongest relationships are built on two things:
 → High expectations
 → High support
One without the other breaks people.

If you expect a lot but don't show up – you create pressure, fear, and disconnection.
If you support a lot but don’t expect much – you hold people back without meaning to. (Some moms unintentionally do that)

People grow when they face high expectations and receive high support at the same time.

I want to be the kind of person who says:
"I believe in what you could become – and I’ll be here while you try"
My style is less "sink or swim" and more like: "Let’s build the damn boat together"

***
To everyone out there supporting their teammates and challenging them to go further: 
You’re doing the real work.
Keep going. 💪

You probably know the old story called "The Stone Soup".
A pilgrim comes to an old woman and asks her for some soup.

"There’s nothing to cook with", – she said.

"No problem", the pilgrim says.
"We’ll make it from a stone. Just add some water."

Then, little by little:
"Maybe a handful of grain?"
"And a bit of butter."
"Just a pinch of salt."
"And, well, a little sugar."

In the end, there’s a great soup.

"And when shall we eat the stone?", asks an old woman.
"The stone... hm", – pilgrim hesitates. "Well, the stone didn't cook well this time, we'll eat it later".

That’s pretty much how AI works these days.

The AI Stone Soup

AI/ML teams inside companies needs less of this:

- "we’ll dump all our files and AI will answer our questions"
- projects where nobody knows what workflow we want to fix, but everyone thinks a chat will solve it
- "let’s ask LLM to predict the future from our messy data"
- trying to use a chat model for a clear classification task
- hoping a bot will fix messy data instead of fixing the data itself

And more of this:
- one clear use-case that saves time or money right now
- clean data flows: what comes in, what goes out, who owns it
- classic ML where it fits better than a chat model
- metrics: what we speed up, what we reduce, what we improve
- steady weekly progress, not “I launched it and now I expect 100% quality right away” 

If companies focused on this shift, half of the "AI failures" would disappear.

Co-founder @ Epoch8 AI/ML agency – We help founders and internal entrepreneurs launch real products built on computer vision, data, and AI ❤️ eCommerce & retail, warehousing & logistics, LegalTech, HealthTech, startups

taplio

Olga Tatarinova's Linkedin Analytics

Olga Tatarinova

Performance ScoreOlga Tatarinova's LinkedIn Performance

What is Olga talking about?

Olga Tatarinova's Best Posts (last 30 days)

Ready to grow your LinkedIn presence?

taplio

.css-6ytdur{background-image:linear-gradient(284deg, rgba(39, 43, 72, 1) 0%, rgba(33, 47, 183, 1) 100%);color:transparent;-webkit-background-clip:text;background-clip:text;}Olga Tatarinova's Linkedin Analytics

Olga Tatarinova

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);}Olga Tatarinova's LinkedIn Performance

What is Olga talking about?

Olga Tatarinova's Best Posts .css-1emoo84{font-size:var(--chakra-fontSizes-md);font-weight:400;color:var(--chakra-colors-gray-600);}(last 30 days)

Ready to grow your LinkedIn presence?

Olga Tatarinova's Linkedin Analytics

Performance ScoreOlga Tatarinova's LinkedIn Performance

Olga Tatarinova's Best Posts (last 30 days)