About
I teach how people and organizations can thrive in the age of AI. My work focuses on…
Services
Articles by Kirill
Activity
63K followers
-
Kirill Eremenko shared thisEvery enterprise has a 20-year backlog of bad data. Twenty years. Sitting there. Everyone knows it's a problem. Nobody can fix it because the manual effort required is just... brutal. Rohit Choudhary has been staring at this problem for over a decade. He coined the term "data observability" back in 2018 — before most people were even thinking about it. He's built three companies with the same co-founder over 14 years. And the thing he keeps coming back to is this: The manual era of data management is over. Here's the example that made it click for me. Imagine you're a CMO trying to improve targeting accuracy across US zip codes. Your zip code data comes from dozens of sources — internal, external, third-party. It's never accurate. And you're dealing with petabytes of it. In the old world, you'd assign a team to find the critical data elements, write quality rules, trace the pipeline, fix what's broken. Weeks. Months. And by the time you're done, something upstream has already changed. Now imagine an AI agent that sifts through all that metadata, takes you straight to the broken data assets, tells you what's wrong, and tells you how to fix it. That's not a future scenario. That's happening now. And here's the part that really shifts your thinking... Rohit says fixing bad data at the point where someone's already trying to use it costs roughly a thousand times more than catching it as it flows through the pipeline. A thousand times. He compared it to the difference between braking on the highway versus crashing — one is a minor inconvenience, the other is cops, insurance, and paralysis. Most data engineering teams today spend 70-75% of their effort not on innovation, not on building the next thing — but on cleaning up old messes. AI agents are about to flip that ratio. The backlog that couldn't be touched for two decades? It's about to get cleared in two years. Kirill P.S. Rohit Choudhary — founder and CEO of Acceldata, the company that coined "data observability" — shared this on our latest SuperDataScience Podcast episode. Link in the comments.
-
Kirill Eremenko shared thisI got my Uber Eats order today. Looked at the receipt. Next to my name — a smiley face. And the letters "AI." It's probably just a staff code. A coincidence. But it made me stop and smile. And think about everyone learning AI. ...in the morning before work ...during your lunch break ...after the day is done and the kids are off to bed To every student who has ever watched one of my videos, taken one of my courses, or read one of these posts — thank you. Genuinely. You are the reason I keep going. 10 years ago I started sharing what I was learning about data and AI. I'm not a researcher. I don't have all the answers. I just found something I loved and decided to share it — openly, consistently, without waiting until I was "expert enough." And moments like this remind me why that decision changed my life. Don't wait until you know enough to share. You never will. Share what you're learning right now. Post about it. Talk about it. Put it out into the world. You have no idea whose life it will touch — or where it will show up. Maybe on a Zambrero receipt one Tuesday afternoon 🤷♂️
-
Kirill Eremenko shared thisI had a call with one of our cohort members this week. Mid-level backend developer. Java, distributed systems, a few years of solid experience. Smart guy. He told me he almost didn't sign up. Not because of time. Not because of money. Because he thought building with AI meant building the AI. Like ... training models. Architecting neural networks. Reproducing what OpenAI or Anthropic spent billions creating. "It felt overwhelming," he said. "The responses feel almost human. I didn't think I could build something like that." **Here's the gap in thinking most developers have:** You don't build the AI. You call it. Same way you call a payment API. Same way you call a weather API. You pass in inputs, you get back outputs, and you build something useful on top. That's it. Once he understood that, everything changed. He went from "this isn't for me" to "okay, I can actually do this." And now he's in week four of a six-week program - building a working AI project, on schedule. The barrier to AI engineering isn't intelligence. It's a mental model. Most developers are already 80% of the way there. They know how to write code. They know how to work with APIs. They know how to think through systems. They just need to realize that AI is the API - and they already know how to use those. Build on top, Kirill P.S. The other thing he said stuck with me: his colleagues are using AI only to write code faster. He wanted to understand what it can actually do. That curiosity is the real differentiator.
-
Kirill Eremenko shared thisThis month Anthropic dropped a new study. This is not just "theory". It's real usage data from millions of professional conversations with AI. Anthropic mapped every occupational category twice. Once for what AI can theoretically do. Once for what it's actually doing. The chart tells the whole story. Blue = what AI could theoretically handle. Red = what's actually happening today. The gap is enormous. Computer programmers are the most exposed occupation at 75% coverage. Customer service reps hit 70%. Data entry at 67%. Meanwhile — cooks, bartenders, mechanics, construction workers? Barely any AI exposure at all. Here's what most people get wrong about this data. They see the gap and think "oh good, AI isn't coming for my job yet." I see the gap and think: somebody has to close it. Companies know AI can transform their operations. They just don't have the people who can actually make it happen. The technology is there, but they still need people who can actually make it happen. Think about that for a second... BUSINESSES ARE IN A RACE TO CLOSE THE GAP And it means there's massive demand building — right now — for professionals who can bridge the gap between what AI can do and what businesses are actually doing with it. You don't need a PhD. You don't need to be a researcher. You need to understand your domain and know how to apply AI inside it. The finance person who can automate reporting workflows with AI? Incredibly valuable. The legal professional who builds AI-assisted research pipelines? Every firm wants them. The best career moves right now aren't about competing with AI. They're about becoming the person who makes AI useful in your field. The study found no increase in unemployment for AI-exposed workers. The disruption hasn't arrived yet. But the opportunity is already here. That almost never happens. You rarely get a head start like this. Use it. What's your field — and how are you using AI in it? 👇
-
Kirill Eremenko shared thisIn 2013, Kyunghyun Cho arrived at Yoshua Bengio's lab in Montreal knowing nothing about machine translation. Bengio had given him four research options. Cho picked the one that scared him most. The lab was tiny back then, about 35 people total. Open office, desks everywhere. Bengio kept sending every visiting intern Cho's way because he had this deep conviction that machine translation was the next big problem to crack. One of those interns was Dima Bahdanau. One morning, Cho walked into the lab with an idea he'd been mulling overnight. Before he could say anything, Dima spoke up. "I think I had some nice idea." Cho let him go first. Dima started sketching it out on paper. Step by step. A mechanism that would let the model look back at specific words in the original sentence while generating a translation, selectively focusing on what mattered most at each moment. Halfway through the explanation... Cho knew. He immediately sensed it was going to work. It was one of the only ideas in his entire career where that happened. His own overnight idea was so weak by comparison that he never even said it out loud. Dima coded it up in three to four days. It worked exactly as they'd hoped. That mechanism was called "attention." It became the foundation of the Transformer. Which became the foundation of GPT, Claude, Gemini — essentially everything powering AI today. A visiting intern. A sketch on paper. A morning in Montreal. The paper has been cited over 40,000 times. Cho's total citations sit around 200,000. But the moment that mattered most? A quiet conversation where one person let the other person go first. Kirill P.S. KyungHyun Cho — co-author of the original attention paper, Professor at NYU, co-director of the Global AI Frontier Lab — shared this on our latest SuperDataScience Podcast episode. Link in the comments.
-
Kirill Eremenko shared thisLOVED this story. The power of getting out there and spreading your message. Applies equally to companies and individuals -- don't be afraid to be noticed. Especially when you have great things to share! It brings opportunities you weren't even expecting. Put yourself out there, KirillKirill Eremenko shared thisDon’t underestimate the power of the podcast 🎙️ They can have a lot more impact than you might think, especially if you’re a founder. Towards the end of last year, I was a guest on Jon Krohn’s Super Data Science podcast – which has 194,000 subscribers – and I’m glad I did. Midway through the conversation, Jon brought up an MIT study that found 95% of AI projects never deliver value in production. He wanted to know how we stay in the 5%. My answer? We refuse to start a project without first agreeing what "good" looks like – the KPIs, the ROI, the metrics. Slightly boring, maybe. But it's why it works. The episode has racked up a respectable 19,228 views on Jon’s YouTube channel – but it was what happened off the back of the pod going live that made it so worthwhile for me and the team. We had zero expectations, just saw it as an opportunity to engage in a fun conversation and get our name out there a bit. However, we ended up hiring amazing new team members (looking at you Sean Thompson and Rachael Holtzberg 👀) and also saw brand new customer projects go to the next level (more on that later). All of them had listened to the podcast, or heard about us off the back of the podcast. We're always keen to get into the big tech publications, but it just shows how powerful new media can be too. Have you found any other fun ways of getting your name out there? Have you got any interesting work or opps out of a quick podcast sesh? Link to the full Super Data Science podcast episode in the comments!
-
Kirill Eremenko shared thisOn March 5th, Amazon's North American marketplaces went down. 6.3 million orders lost. Amazon engineer: "People are becoming so reliant on AI that essentially they stop reviewing the code altogether." Connect the dots? Everyone's vibe coding now. Prompt in. ➡️ Code out. ➡️ Ship it. ➡️ Next feature. And honestly? For getting something working fast, it's remarkable. Cursor, Claude, Copilot — the tools are genuinely good. A working prototype in an afternoon isn't hype anymore… it's Tuesday. But here's what's also happening: Engineers are shipping code they can't read ... Fixing errors they can't interpret ... Building systems they fundamentally don't understand ... … and only discovering that when something breaks in production. Vibe coding doesn't give you intuition. It borrows someone else's. Short-Term: that's fine when everything works. It's a crisis when it doesn't. The picture is much worse Long-Term: Every shortcut you don't question is a muscle you stop using. A year from now, you've shipped a lot — but you know less than when you started. The engineers who actually thrive with these tools? They never stop asking why. - They read the code. - They rewrite pieces by hand just to make sure they still can. - They vibe code, while learning & staying sharp. - All. The. Time. And they know when the AI is wrong. They can feel it. That's the gap vibe coding can't close for you. Use the tools. Ship fast. But spend at least 20% of your time going one level deeper — why did that model choice make sense? * What is that API call actually doing? * How would you rebuild this if the AI wasn't there? That's not extra work… that's the work. As promised, value in under a minute. Stay sharp, Kirill P.S. You DON’T need to review every line of the generated code. But you DO need to understand what each section of the code is doing and why it’s there. The engineers who scare me aren't the ones who can't code. They're the ones who can ship anything and explain nothing.
-
Kirill Eremenko shared thisCohort 5 of the AI Engineer Sprint is in 24 hours. 3 spots left. Build this: Week 1 — Your First LLM (API calls, prompts, conversation history) Week 2 — Your First AI App (RAG, guardrails, dynamic context) Week 3 — Tool Calling (LLM tool handling, multi-tool orchestration) Week 4 — RAG Pipeline (vectors, chunking, embeddings, ChromaDB) Week 5 — Deploy to Hugging Face Spaces (go live, debug in production) Week 6 — Agentic AI (agentic loops, multi-step execution, orchestration) The photo is from Cohorts 1, 2, and 3. These people came in at completely different levels. Some had years of software engineering experience. Some had never made an API call. Six weeks later, they're building RAG pipelines, deploying Digital Twins, and writing agentic AI systems. My team and I are there with you every step, every week. Helping you learn AI Engineering and keeping you accountable to your goals. 3 spots left. Kicking off call is in 24 hours! Interested? 👉 Reply below and I'll DM you the details.
-
Kirill Eremenko shared thisI built an AI-powered Digital Twin of myself. It answers questions about my background, retrieves relevant info using RAG, and notifies me when someone wants to connect. 👉 Try it here: https://lnkd.in/gHqZhD_A Took me 5 weeks of evening sessions to go from zero API calls to a deployed app. Under the hood: - Prompt engineering (system vs user prompts) - Tokenization and API Cost management - Conversation history & context management - Building chat UIs with Gradio - RAG (chunking, embeddings, vector stores) - LLM tool calling (parallel & sequential calls) - Deploying to Hugging Face Spaces If you want to get hands-on with AI engineering, join the next AI Engineer Sprint here: https://lnkd.in/gSTVbcsH #AIEngineerSprint #SuperDataScience
Kirill Eremenko commented on a post
5h
Full episode: https://www.superdatascience.com/979
This episode is brought to you by Cisco and by Open Data Science Conference (ODSC).
Experience & Education
-
SuperDataScience
View Kirill’s full experience
See their title, tenure and more.
or
Licenses & Certifications
Volunteer Experience
-
Guest Lecturer
The University of Queensland
- 5 months
Education
Several guest lectures on topics including direct marketing, segmentation, and customer/market analytics. These lectures have proven valuable for shaping Kirill's presentation skills as well as connecting students with real world analytics and developing a recruitment channel.
Publications
-
Stability of the neodymium laser with an unstable resonator
52nd open academic conference of the Moscow Institute of Physics and Technology
Honors & Awards
-
Dean's commendation award
University of Queensland
-
Honors
Moscow Institute of Physics and Technology
-
Honors
ZFTSH Correspondence School
-
Junior Chess Champion of Zimbabwe
Zimbabwean Chess Federation
-
GMAT
Graduate Management Admission Council
740 score in GMAT (97th percentile)
-
IELTS
Cambridge English Language Assessment
IELTS score 9.0 (max 9.0)
Languages
-
English
Native or bilingual proficiency
-
Russian
Native or bilingual proficiency
-
French
Elementary proficiency
-
Persian
Elementary proficiency
Organizations
-
Institute of Analytics Professionals Australia (IAPA)
Member
- Present
Other similar profiles
Explore more posts
Explore collaborative articles
We’re unlocking community knowledge in a new way. Experts add insights directly into each article, started with the help of AI.
Explore More