About
O’Reilly Author | Tech & Product Executive
- Global Scale: 25+ years of…
Articles by Irakli
-
Microservice Architecture – Avoiding Coordination Costs
Microservice Architecture – Avoiding Coordination Costs
One of the core themes of the Microservice Architecture book that my co-authors and I published with O'Reilly recently…
1 Comment
Activity
3K followers
-
Irakli Nadareishvili shared thisNo new version of an LLM will be a true breakthrough, until we fix memory handling! This is your regularly scheduled PSA that the biggest outstanding challenge in making LLMs think more like humans, and becoming true collaborators, is that their memory and learning work very differently from ours. Humans learn bit by bit, from every interaction with the real world, not just from curated training. Touch a hot stove once, remember the pain forever. The lesson gets baked into how you reason about the world. LLMs don't do this. Weights are frozen at inference. You correct the model in today's chat session, in the future it very likely makes the same mistake, and will definitely make the same mistake with another user, in another session. This happens, because there is no standard, architectural path from "user feedback" to "updated priors." In-context memory evaporates over time, and is mostly user-specific, anyway. The current workaround — tools, retrieval, skills, longer context — is externalized lookup. A better filing cabinet, not a richer mental model. The priors don't move. Until incremental learning is solved at the architecture level, continued scaling and more tooling will give us more capable assistants. It won't give us the breakthrough. Sorry to be the bearer of this news, but until you see "incremental learning going directly into LLM memory" in release notes, no new version announcement is revolutionary, they are refinements + marketing
-
Irakli Nadareishvili shared thisSometimes you read a post (yes, even on LinkedIn) that resonates so much, that you keep nodding to it and smiling at it, as you read it. This is one of those and an important one, at that. Thank you, Joe.Irakli Nadareishvili shared thisAdmiral Grace Hopper invented the compiler in 1952. The response from her colleagues was that computers couldn’t understand English, only mathematics, and she was wasting her time. She kept a clock on her desk that ran backwards. It worked perfectly well. She used it to remind people that the way things had always been done was a choice, not a law. The pattern in computing’s first decade is the one worth studying right now. The people who moved the field forward were not the ones with the cleanest theory. They were the ones who built something, watched it fail, and built it again before the committee finished meeting. Hopper shipped COBOL while the industry was still arguing about whether high-level languages were a serious idea. By the time the argument resolved, the answer was already in production. Most enterprise AI programs in 2026 look nothing like that. They look like a steering committee, a vendor evaluation, a center of excellence, and a pilot that has been running for eleven months. Meanwhile a smaller group of operators has rebuilt entire workflows around agents this year. We have working systems handling real volume, a clear view of where the models break, and the institutional muscle that comes from operating AI in production. While the committee is still picking a framework. Hopper did not optimize the punch card era. She ended it. The work in front of enterprise leaders right now is the same. AI native is not a layer on top of the existing org. It is the org, redesigned around what the technology actually makes possible. Grace Hopper. Rear admiral, software pioneer, professional troublemaker. Here early 1970s in front of a PDP/11.
-
Irakli Nadareishvili shared thisThere are some old habits and beliefs about software quality, that we will need to rethink in the age of GenAI. The ability of coding agents to accelerate technical teams is undeniable. What is also clear is the need to keep it safe. Hallucinations, messy code, code that is too large, and faked tests are all real risks that teams must address and mitigate. Personally, I don't think ANY of these are new problems; we've seen messy code, fake tests, and misunderstood requirements in work produced by human engineers. What is new is scale! Coding agents can go much faster than any human and produce much larger output, so how do we deal with these familiar problems at 10x the scale? In a world where: solutions are designed with an LLM agent's assistance, coded almost entirely by an agent, reviewed by an agent, and tests also written by an agent, how do we keep safety? And for some of us, who are also subject to compliance, how do we keep it compliant? Let's just say one thing right away: scaling human review of code in the old-fashioned line-by-line approach simply won't work. Attempts to fight "large-PRs-by-Claude" are futile. Thousands of smaller PRs are not necessarily more reviewable than dozens of larger ones. Old ways won't scale; we need new ways. Before we get to even discussing what the new ways are, I believe we need to clearly identify what "quality" means for software going forward. For the longest time, neat, clean, concise code was a hallmark of good code. Now that most code is written by agents, and increasingly will be read by agents, this metric is doomed to fail. I believe we should focus on outcomes instead. What if we focused on non-negotiable traits: 1. Does the code behave correctly? 2. Does the code perform and scale as desired in terms of latency, throughput, memory consumption, etc.? 3. Is the implementation secure? 4. Is the architecture and code evolvable, i.e., easy to modify (not necessarily by humans!)? 5. Is the code reasonably easy to debug when problems occur (not necessarily by humans)? There may be more traits we need to consider, but you get the point. If the generated code passes these non-negotiable quality checks, should we care how long it is, or how "clean" it is? Maybe we very soon won't. Maybe the actual source code written in C, Java, Rust, etc. will become as "irrelevant" as machine code had become in the past. As long as it passes the new quality metrics, it may be good enough. Possibly even "perfect," if such a thing existed :) Something to start thinking about...
-
Irakli Nadareishvili shared thisCool job alert! Rights & Pricing team at Apple is looking for a Quality Automation Engineer to help build scalable automation, drive quality at scale, and leverage AI in testing. Great opportunity to work for an amazing leader Moulaye Keita on high-impact systems across Rights & Pricing - Partner Onboarding, with strong ownership and collaboration across teams. Apply or share: https://lnkd.in/gUn7VVQVQuality Automation Engineer, Rights & Pricing- Partner Onboarding Business - Jobs - Careers at AppleQuality Automation Engineer, Rights & Pricing- Partner Onboarding Business - Jobs - Careers at Apple
-
Irakli Nadareishvili shared thisMy team is hiring! If this role is a good fit, please apply! You will get to work at, in my strong opinion, world's most iconic company, with brilliant humans, solving really interesting problems, using cutting edge AI and non-AI tech https://lnkd.in/gQsgsrTzSr. Software Engineering Manager, Pricing-Subscriptions & Infrastructure - Jobs - Careers at AppleSr. Software Engineering Manager, Pricing-Subscriptions & Infrastructure - Jobs - Careers at Apple
-
Irakli Nadareishvili shared thisSpeed up Claude when working with large codebases! If you work with large codebases (monorepos, legacy systems, or really any large company's codebase), you know the pain of having Claude analyze massive source trees: it's slow and burns through tokens. If that sounds familiar, install tokei, ast-grep, rg, and fd via brew, then add the snippet below to your CLAUDE.md so it uses these tools instead of the slower defaults. Source: https://lnkd.in/gPyhKGNj
-
Irakli Nadareishvili shared thisGenAI is less like the Steam Engine, and more like the Photo Camera. In the tech world, we see some bragging on LinkedIn that "100% of my code is now written by [Agent Blah]." People keep using that phrase, but it doesn't mean what most think it does. Indeed, 100% of the pixels in a photograph are "machine-created," not hand-painted. Does that mean there’s no difference between an amateur snapping a forgettable selfie and a professional photographer capturing a masterpiece? Of course not. When navigating massive shifts, we need the right point of reference. And I’m not a fan of comparing GenAI to the Industrial Revolution. The parallels are thin and, frankly, not very helpful. Instead, the analogy I use, is the rise of mass photography, in the early 20th century. When photo cameras arrived, they "democratized" capturing of visual reality. It took what used to be a grueling, manual skill, painting a realistic scene, and made it accessible to everyone. But it didn't make everyone an artist. Instead, it gave birth to a massive, lucrative new occupation: the professional photographer. Interestingly, it also never "nullified" the artist with a palette and a brush. Photography just forced painting to evolve deeper into something more expressive and less about pure documentation. When cameras were invented, the mechanics of capturing light became trivial. In our world, translating intent into code is becoming trivial. But that only places more emphasis on the things that aren't trivial: * Creative taste. * Identifying the right problem to solve. * Knowing the "angle," the "frame," and the "light." * Building proper technical architecture, security, and delivering winning customer empathy. The Creative Bottleneck is Gone By removing the bottleneck of mastering pure mechanics, we’ve opened the door for more people to build. I’d wager there are more professional photographers today than there were "great painters" in the 18th century. AI will likely do the same for software and engineering. Think of Claude or any LLM as a high-end photo camera. It’s powerful, it’s "smart," but it isn't fully automated, or replacing humans. Not yet, anyway. "I’m scared for the future. What is the world turning into?" If you haven’t heard some version of this lately regarding AI and machine learning, I envy you. With GenAI, it’s the best time in history to be a creator, yet there’s a constant, nagging fear: What’s next? Is AI becoming too powerful? I genuinely have no idea what can happen in the next 5–10 years. I am certain nobody does. Anyone who claims to know exactly what the world will look like, long-term, is probably a charlatan. However, I have a very strong opinion on the "right here, right now." This generation of AI tools are not as scary as the headlines suggest. Use the good parts of it, and let’s also work together to ensure the future AI doesn't turn into major evil.
-
Irakli Nadareishvili shared thisWe need to have an honest conversation about PRs (Pull Requests) and code change safety in the new, Claude-driven world. PRs were never foolproof. There, I said it. Even when it was just humans reviewing humans' code, things slipped through. But in a world where Claude is producing code at an order of magnitude higher volumes , the traditional PR process isn’t just weak, it’s broken. The "Claude writes, humans review everything" policy is well-intended, but it’s naive. Humans simply cannot keep up with the volume. And let’s be honest: there are serious questions to be asked when we have an LLM reviewing code that another LLM just wrote. I’m not saying PRs are dead. But right now, most teams are just "sprinkling AI" on top of 20-year-old processes. It’s walking by default when you finally have wings. We don't need faster manual reviews; we need AI-native reviews. To me, an AI-native PR looks like this: - The Prompt Log: Access to the "plan-mode" conversation between agent and a human, and prompts that led to the changeset. These logs are often more telling than seeing the long code listing. - The Safety Net: A clear synopsis of the automated tests (new and existing) that actually cover the changed logic. - The "Second Opinion": A summary from a separate AI agent session, to provide LLM-backed line-by-line analysis that humans struggle with. This structure is far more telling than a human squinting at thousands of lines of generated code. It shifts the human’s role from "line-by-line debugger" to "architectural judge." It’s a shift worth trying. Give it a shot, and please share your experience!
-
Irakli Nadareishvili shared this"Claude Code Only Works on New Code" You’ve likely heard it before: developers getting amazing results when asking agents like Claude to build something from scratch, but having comparatively less success using it for "actual work", you know, that 10-plus-years-old Java monolith everything actually runs on. It doesn't have to be this way. The core issue is context. When you build something new, Claude is there from the first prompt. It understands every design decision because it made them. But when you drop Claude into a legacy codebase, that contextual knowledge is gone. Even worse, if you try to "fix" this by shoehorning your entire architecture into a single, top-level CLAUDE.md file, you’ll likely hit a wall, once that file exceeds ~200 lines, instruction reliability craters. The solution? Don't document your codebase all in one place! Instead, leave a trail of "breadcrumb" CLAUDE.md files throughout your directory structure. By scoping context files to specific modules or packages, they are only loaded when Claude is actually working in that section of the code. Here is a prompt to help Claude automate this organization for you: --- "Document everything you know about this project so that a new session of Claude Code can be productive immediately. Analyze and survey the entire project first to understand the architecture, purpose, and README. Create fine-grained context files scoped to the specific directories where they matter. Keep the main CLAUDE.md minimal. Include only a single-paragraph synopsis of the project goal. Define slash commands and skills for repeatable tasks to ensure new sessions follow established architectural and stylistic patterns. CRITICAL: Present a plan of the changes you intend to make before modifying any files." --- I also created a command version of this prompt, for usability. It has additional instructions to force Claude in creating skills (LI post limits wouldn't allow to show entire prompt). You can find this command, along with a helpful hook, in a cleaner format on GitHub here: https://lnkd.in/gGk6hR4Y
-
Irakli Nadareishvili reacted on thisIrakli Nadareishvili reacted on thisPoint72's big hire from JPMorgan. Steve Cohen's hedge fund Point72 has hired JPMorgan's head of electronic trading services Michael Heller. Michael has joined Point72 as the head of systematic trading infrastructure in New York after spending more than four years at JPMorgan. Previously, he served as the global head of electronic trading platforms at Goldman Sachs for more than 16 years. Here are some of the recent moves at Bank of America, Morgan Stanley, Citi and Man Group: https://lnkd.in/eEmw_Nc5
-
Irakli Nadareishvili reacted on thisIrakli Nadareishvili reacted on thisAfter 13.5 incredible years at JPMorgan Chase, I’ve made the decision to move on and begin a new chapter. I’m excited to share that I’ve accepted a Program Manager position with Huntington National Bank. Chase has been more than just a workplace for me—it’s where I grew my career while raising my three children. I’m deeply grateful for the opportunities I’ve had along the way, serving as an Incident Manager, Operational Specialist, Project Manager, Portfolio Manager, Agility Lead, PAL, and most recently, leading transformation efforts across 60+ teams. What I’ll miss most are the relationships—the coworkers and friends who made this journey so meaningful. Thank you to everyone who has been part of my story over the past 13.5 years. I’m proud of what we accomplished together and look forward to what’s ahead. #Gratitude #Leadership #Transformation #NextChapter #HuntingtonNationalBank
-
Irakli Nadareishvili reacted on thisIrakli Nadareishvili reacted on thisIf you know, you know: for Apple employees, getting your picture in front of the 1 Infinite Loop sign is basically a rite of passage. And I finally did it! I am cheesing way too hard and not even a little sorry about it.
-
Irakli Nadareishvili reacted on thisIrakli Nadareishvili reacted on thisAdmiral Grace Hopper invented the compiler in 1952. The response from her colleagues was that computers couldn’t understand English, only mathematics, and she was wasting her time. She kept a clock on her desk that ran backwards. It worked perfectly well. She used it to remind people that the way things had always been done was a choice, not a law. The pattern in computing’s first decade is the one worth studying right now. The people who moved the field forward were not the ones with the cleanest theory. They were the ones who built something, watched it fail, and built it again before the committee finished meeting. Hopper shipped COBOL while the industry was still arguing about whether high-level languages were a serious idea. By the time the argument resolved, the answer was already in production. Most enterprise AI programs in 2026 look nothing like that. They look like a steering committee, a vendor evaluation, a center of excellence, and a pilot that has been running for eleven months. Meanwhile a smaller group of operators has rebuilt entire workflows around agents this year. We have working systems handling real volume, a clear view of where the models break, and the institutional muscle that comes from operating AI in production. While the committee is still picking a framework. Hopper did not optimize the punch card era. She ended it. The work in front of enterprise leaders right now is the same. AI native is not a layer on top of the existing org. It is the org, redesigned around what the technology actually makes possible. Grace Hopper. Rear admiral, software pioneer, professional troublemaker. Here early 1970s in front of a PDP/11.
-
Irakli Nadareishvili liked thisCome work with us and join an awesome team 🖤:
Experience
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
Education
Publications
-
Microservice Architecture
O'Reilly Media
Microservices can have a positive impact on your enterprise—just ask Amazon or Netflix—but you can fall into many traps if you don’t approach them in the right way. This practical guide covers the entire microservices landscape, including the principles, technologies, and methodologies of the unique, modular style of system building. You’ll learn about the experiences of organizations around the globe that have successfully adopted microservices.
In three parts, this book explains how…Microservices can have a positive impact on your enterprise—just ask Amazon or Netflix—but you can fall into many traps if you don’t approach them in the right way. This practical guide covers the entire microservices landscape, including the principles, technologies, and methodologies of the unique, modular style of system building. You’ll learn about the experiences of organizations around the globe that have successfully adopted microservices.
In three parts, this book explains how the services work and what it means to build an application the Microservices Way. You’ll explore a design-based approach to microservice architecture with guidance for implementing various elements. And you’ll get a set of recipes and practices for meeting practical, organizational, and cultural challenges to microservice adoption.Other authors -
-
Microservices shift complexity to where it belongs
O'Reilly Radar
See publicationMicroservices are no free lunch, but they are smart nutrition.
-
Two-Pulse Spin Echo In Two-Level Systems Inside Amorphous Ferromagnets
Acta Physica Slovaca
Patents
-
Object-Level Encryption
Issued US 2023/0053394 A1
A method, apparatus, or computer-readable medium for performing object-level encryption and key rotations
Other inventors -
-
Generation of Format-Preserving Tokens to Secure Sensitive Data
Filed 17/946,710
Other inventors -
Honors & Awards
-
Health 2.0.: Best Provider Facing Design Winner
Health 2.0 – 2015 Design for Health Awards
http://www.health2con.com/news/2015/04/09/2015-design-for-health-awards-winners-announced/ (Company was still called Urgent Consult, back then)
Languages
-
English
Full professional proficiency
-
Georgian
Native or bilingual proficiency
-
Russian
Full professional proficiency
Recommendations received
12 people have recommended Irakli
Join now to viewOther similar profiles
Explore more posts
-
Rhonda Coleman Albazie
The Larger California Urban Divide The SF vs Southern CA contrast is part of a broader divide emerging in California: Two models of cities are evolving: Dense progressive tech metros • San Francisco • parts of Los Angeles • Seattle vs. Affluent coastal suburban enclaves • North County San Diego • Orange County • parts of Santa Barbara County These places operate under very different political cultures, zoning rules, and urban planning philosophies.
-
Ken Zhang
Today marks the final day of NVIDIA GTC in Silicon Valley, an event that has experienced significant growth. Downtown San Jose is bustling with participants, highlighting the pervasive influence of AI as it disrupts various industries. EdgeNectar is at the forefront, pioneering AI for private 5G zero-touch operation. This plug-and-play solution enables enterprise AI applications with low latency, providing reliable and secure connectivity.
2 Comments -
Derek Kerton
We chose a good subject for our march Autotech Council meeting, and after this year's CES, it seems even more obvious. Our meeting on the AI Defined Vehicle will show how it's a natural next step after the Software Defined Vehicle phase that we've been in Tesla launched the Model S in 2012. Now, most carmakers have SDVs. It's time to layer on the AI in user-facing roles, and in invisible functions managing the vehicle.
-
Brett A. Hurt
Adding to Bob Pearson’s comment, I would also emphasize that the biggest advantage is capital availability and the risk-taking mindset that goes along with it. Silicon Valley VC was being born by people like Arthur Patterson (one of my longest-time backers and mentors), since I was born. Austin, by contrast, was the size of Anchorage, Alaska when I was born. Austin is definitely doing the best it ever has and I’m happy to be here. But the biggest risk-taking and capital-availability advantage remains in SV due to the compounding effect over decades. Also, remember Debra and I lived in SF during the massive boom and bust years, from the beginning of 2000 to mid 2003. There will certainly be an AI application bust, like Mitchell Green, is predicting but overall the trajectory is correct. AI will affect everything and therefore deserves the capital allocation it’s receiving. But most AI applications will not be durable enough to survive the growth of frontier model scope and ambitions, similar to Microsoft Windows back in the day and the many companies it put out of business.
-
Xiaowei Jiang
Looking forward to speaking at AI4 2026 (Aug 4–6, Las Vegas). AI discussions often focus on models. In production systems, the harder problem is decisions — multiple services and agents acting on slightly different state, leading to conflicting outcomes. I’ll talk about the operational context layer needed so automated decisions can actually be trusted in workflows like payments, trading, and security. Hope to see you there.
1 Comment -
Tamara Pettersson
🎓 Empowering Groundbreaking Research at UC Berkeley 🎓 The University of California, Berkeley, is at the forefront of innovation - conducting research that generates petabytes of data across genomics, astronomy, social sciences, and more. But managing and storing this data efficiently is no small task. 🌟 Enter Cloudian. In 2024, Berkeley IT revolutionized its storage approach by offering researchers Cloudian’s scalable, cost-effective data storage solution. At just $48 per terabyte per year (for five years), this platform provides: 📌 Massive scalability – Seamlessly grow to petabytes of storage 📌 Fast, reliable access – Data available anytime, anywhere 📌 Cost savings – No hidden fees, predictable budgeting 📌 Flexibility – Support for all data types, from hot to archived With Cloudian, Berkeley researchers 🔬 can focus on discovery, not data logistics. Curious how Berkeley tackled the challenge of growing research data? Check out the full story here: https://bit.ly/4h4mGtI #DataStorage #Cloudian #UCberkeley #DataScalability #PetabyteScale #ObjectStorage #DataResearch #StorageInfrastrcture #GenomicStorage
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content