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.

🚀 "How to teach AI assistants to reason – not to guess"

Last month, I spoke at SmartData 2025 about something we’ve been wrestling with for over a year.

Everyone is building RAG systems – assistants that fetch text and generate answers. But even with perfect retrieval, they still hallucinate.

Why? Because language models don’t understand how your domain works – they just imitate its surface.

So we spent a year asking:
Can we build an AI assistant that actually understands how your business domain works – and reasons instead of guessing?

The result is what we call Semantic RAG – an approach that:
 🧩 separates domain knowledge from the model,
 🔍 connects both structured and unstructured data,
 🧠 and gives the assistant a playbook to reason and decide how to answer each question.

Semantic RAG Is not just another RAG tweak – it’s a way to teach AI to reason like your business experts do, through *data modeling*.

Huge thanks to the SmartData team for the energy – and to everyone who came to chat after the talk. And huge thanks to Alexey Makhotkin, creator of the Minimal Modeling approach, for the never-ending inspiration.

Slides are below 👇

Olga Tatarinova

Co-founder @ Epoch8 – AI/ML/engineering agency for high-load and data-intensive systems | Sharing lessons from 200+ projects

Semantic RAG – how to teach AI assistants to reason

The AI Stone Soup

A team of consultants walked into a company.
Processes were tangled, reports delayed, everyone waiting for a silver bullet.

The consultants smiled:
"We’ll implement AI and automate everything!"

The team was impressed. Finally – salvation.

Then the consultants said,
"We just need one small thing… your data."

The managers hesitated, then shared a few Excel files.

"Wonderful. Now, could someone write down how this process actually works?"
Operations added their diagrams.

"Excellent. Maybe a small Python script?"
The engineers pulled one from an old repo.

"Perfect. And some testing, to be sure it really works?"
QA joined in.

"And a bit of DevOps, to keep it alive?"
Ops brought their tools.

Soon the pot was boiling with contributions. Dashboards were humming, reports arrived on time, the process finally ran smooth.

The CEO tasted the result:
"Amazing. This AI really changed everything!"

The consultants just exchanged a glance.

Because the secret was simple:
there was no AI in the pot at all –
just good old process documentation, a few Python scripts,
and people finally talking to each other.

The best startup advice I’ve heard this month

I recently spoke with Kirill Danchenko, co-founder of Readymag.
He’s now building a new startup, this time without VC funds, raising micro-investments from friends and network.

The whole Kirill's startup philosophy sounds fresh to me, especially these two principles he shared:

1. Test ideas not by external response, but by internal resonance

The usual startup playbook says:
find a market → identify a problem → talk to users → build an MVP.

But if it’s not your problem, you test it half-heartedly.
You listen to users but don’t really get them.
You build something but without fire.
And when it doesn’t work, you say, "The idea failed."

Kirill is a contrarian here.
He tests hypotheses by asking:

> "Does this idea excite me enough to live with it for the next 3-5 years?"

If yes - he commits fully.
If no - he lets it go.

Because - as we all witnessed many times - startups don't die from lack of money or temporary operational problems.
They die when the founder loses motivation.

2. Replace big investors with micro-investors

Kirill has already lived through investor dependency -
when someone pulls out and the team suddenly has no salaries and no runway.

Now he’s not going for VCs. 
Instead he raises $3K-$5K micro-investments from friends, family, and peers -
with a promise of a profit share once the startup becomes sustainable.

It’s part Kickstarter, part community building. People take a small risk to be part of something meaningful.
And maybe turn that $3K into $9K along the way.

In the end, it’s a cool combo:
inner drive + outer freedom without investor pressure, but with long-term founder interest - the only fuel that really lasts.

Curious what you think - could this micro-investor model actually work long-term?

PMs – how do you actually make sense of user feedback?

Every PM says "we listen to our users", but if you’ve got thousands of open-text reviews… how do you analyze it?

We recently automated our feedback analysis with GPT – it translates, detects sentiment, categorizes reviews, and even creates weekly summaries ("Top 3 complaints this week", "Website sentiment recovered +20%", etc).

It helped us see, for example, how a website redesign suddenly spiked negative reviews and which exact topics caused it.

But I’m wondering – what's your approach to turning messy review text into something actionable? 

Do you tag manually or automate somehow?

Any frameworks for structuring feedback (categories, topics, summaries)?

What’s your go-to dashboard or visualization?

Always love seeing how others turn messy text into something the team can actually act on.

Sharing our approach in the slides below 👇

8 charts to understand your User's Feedback like a king

💡 “I already know what I’m looking for” – and that changes everything.

Everyone wants their own AI assistant™ these days.
And most teams do the same thing:
dump a pile of docs in → get a pile of random answers out → then wonder why “GenAI doesn’t work.” 

But the real problem is not in algorithms and not in how you chunk or rerank documents.
The real problem is that you never decided what you actually need in the first place.

👀 Imagine you’re an agency.
You’ve got hundreds of contracts with clients and vendors.
Sure, you can slice them into chunks and run a vector search (like everyone does).

But when you ask “When does my contract with vendor X expire?” – you’ll probably get noise instead of answers.

Now imagine you know in advance what to extract:
- counterparty name
- contract end date
- contract terms
- contract amount

Then you build a tiny, meaningful data model – just few fields, but full of structure.

And suddenly your assistant becomes actually useful.

For me, data modeling = clarity.
It’s saying: “I already know what I’m looking for.”
It’s slow.
But it’s the only thing that makes AI assistants truly smart –
because that’s when they start to understand how your world works.

🧩 Ready-made domain data models are pure gold.
They save weeks of work and turn an assistant from a toy into a tool.

PS. I once modeled the financial structure of my own agency –
a domain I know inside out – and it still took two months.
Now imagine doing that for every new industry.
That’s where the real GenAI pain begins. 😅

I often say that AI assistants start to understand better when they know how your business works – what kinds of things exist and how they relate.

But then comes the obvious question:
How do you actually explain it to an AI?

We use Minimal Modeling, a lightweight data modeling method we love, created by Alexey Makhotkin.

It has three concepts: Anchors, Attributes, Links.

⚓ Anchors – the "nouns" of your domain

These are the core entities everything revolves around.
E.g.: User, Product, Order.
Each anchor has its own ID and exists independently.

🏷  Attributes – simple facts about an anchor

Name, description, price, status – anything that describes one thing.
For example, a Product has name, price, specifications.

🔗 Links – relationships between anchors

They show who’s connected to whom.
For example: User –has→ Order, Product –belongs to→ Category, or Employee –reports to→ Manager.
Links literally map the business logic of your system.

Together, anchors + attributes + links let you describe any domain
so that your AI assistant can reason better –
understanding what’s an entity, what’s a fact, and what’s a relationship.

Even if all you have is a pile of unstructured text –
once you tell the model what kind of things it’s looking at,
it already starts reasoning better.

💡 Some practice:
If you’re working with something like a legal contract, you might extract meaningful data into a table.

For example:

⚓ Contract_ID
🏷 Contract date
🏷 Contract Terms & Obligations
⚓ Counterparty_ID
🏷 Counterparty_name

Here we define that we have only two anchors we care about:
⚓ Contract
⚓ Counterparty 

one link:
🔗 Contract – has → Counterparty

and a few attributes.

Most other details in the document – is secondary and less important.

And even if you can’t build an automated extractor –
start small: explain your domain terms in the beginning of the prompt.
Define:
- what "Contract" and "Counterparty" mean 
- how they are related
- what attributes are important

You’ll be surprised how much confusion disappears instantly.

---
So, what would you rather see next?
1️⃣ How to model your data for an AI assistant (a 10-minute walkthrough)
2️⃣ Or a short explainer on LLM context engineering?

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)