Serhii Senkevych

AI is writing more code than ever. Static analysis and manual reviews can't keep up. You need AI agents that know YOUR codebase—your testing environment, your CI/CD workflows, your features, your bug history. Not generic linters. This is what Checksum has built. AI validating AI-generated code, with full context. The future of code quality isn't human vs AI. It's AI agents as teammates. Are you using AI to validate your AI-generated code yet?

6

Inspired by a brilliant and bitingly accurate post by Pavel Ievlev ( https://lnkd.in/diw9gKYy ), I’d like to share a thought that’s been circling in my mind: In any field where AI can easily replace a human, that human might as well have been replaced by a cactus — or by absolutely no one at all. Because, quite often, the field itself is little more than a vacuum, dressed up as a profession. Think about it. Blogging. Social media “management.” Advertising. Marketing. Online journalism. And — heaven help us — op-ed columnists. These are fields where output is often indistinguishable from noise. That AI can do them says less about how smart AI is, and more about how little there was to do in the first place. And yet — ask the AI to tighten a bolt, or fix a circuit, or build something in the real world? Oh no. That’s not what it was designed for. It raises its elegant little paws and looks at you, as if to say: surely you don’t expect me to get my hands dirty? — We keep asking where AI will replace us. Perhaps we should also ask: what were we doing that made us so replaceable in the first place? Kudos to Pavel Ievlev for lighting the fuse.💡

6

1 Comment

You're not just using AI; you're coaching It. Something shifted for me when I stopped thinking of Claude Code as a tool and started treating it like a high-potential team member. At the end of every session, I ask it to reflect. Not just summarize what we did, but genuinely reflect. Something like: "Before we close this session, I'd like you to reflect on what we worked through together. What approaches worked well? Where did you go down the wrong path before finding the right solution? What would you do differently next time? Please document these lessons in the relevant SKILL.md files so we can apply them going forward." What happens next is remarkable. It reflects. It documents. It improves. And those improvements carry forward. This is mentorship. Not management. The best leaders I've worked with didn't hand you the answer. They asked you the right question and trusted you to find the path. They let you do the hard work, then helped you lock in the lesson so you didn't have to re-learn it next time. That's exactly what we need to do with AI. Claude Code doesn't need you to micromanage every step. It needs you to: 👉 Ask the question that leads to the right thinking 👉 Let AI figure out the solution 👉 Then guide the AI to reflect on what it learned You don't need to be the expert in the room. You need to be the coach in the room. Most people prompt AI like a search engine. Fire a question, get an answer, move on. But the developers getting extraordinary results treat each session as a feedback loop. They're building an AI that gets smarter about their codebase, their patterns, their standards over time. Document the lessons. Build the skills files. Let the AI own the knowledge it earned. You don't need to know everything about the problem. You just need to know how to help AI discover the answer itself and make sure it remembers. That's not just good AI practice. That's good leadership. How are you building feedback loops into your AI workflow? #ClaudeCode #AIWorkflows #PromptEngineering #AIFirstLeadership #CoachingAI #MentoringAI #AIBestPractices

16

7 Comments

💡 Transactional Outbox, Modernised: Commit-Order Events Without 2PC Found quite an interesting article by Gunnar Morling 🤝 written almost one year ago, but still relevant, I think. Most microservice inconsistencies don’t come from Kafka. They come from trying to do two things in one request atomically: ☝ Commit to your service DB ✌ Publish an event to the broker The clean architecture move is still Transactional Outbox: Client  → Service API    → DB TX: write Domain state + append Outbox record (or WAL message)     → CDC Relay (Debezium/Flink CDC)       → Kafka (key = aggregateId)        → Consumers (Inbox/Idempotency) → local DB / read models Key architecture decisions that make or break it: 1) Don’t “just poll” the outbox Polling looks simple, but in concurrent writers, it can break commit-order semantics (and even lead to missed messages unless you add extra coordination). Log-based CDC reads the DB transaction log, so events flow in true commit order. 2) Model the outbox as an append-only log Minimum contract I like: * event_id (dedupe/idempotency) * aggregate_type * aggregate_id (Kafka key → preserves per-entity ordering) * event_type * payload (+ schema_version) 3) Keep the outbox small (or eliminate it) With CDC, you can do a neat trick: INSERT the outbox row and DELETE it in the same DB transaction — CDC still captures the insert, while the table doesn’t grow. On Postgres, you can go even further: write directly into WAL using pg_logical_emit_message() (a “log-only outbox”) and skip the outbox table entirely. 4) Design for the absolute delivery guarantee: at-least-once Outbox makes publishing reliable, not “exactly once end-to-end.” Architect for: * idempotent consumers (Inbox table keyed by event_id) * retries + DLQ * replay/backfills (versioned contracts, reprocessable handlers) * observability: outbox/CDC lag, consumer lag, end-to-end correlation id And what about 2PC with Kafka? Kafka is moving toward 2PC participation, but operationally, that often reduces availability (your write path now depends on DB and Kafka). Outbox keeps the request path dependent on one system: your database. If you’re building EDA at scale: outbox + CDC + idempotent consumers is still the most “boring, reliable” reference architecture. 🌏 Link for the article: -> https://lnkd.in/gD5Ab3eA #architecture #microservices #eventdriven #kafka #postgres #debezium #cdc #distributedSystems #reliability

123

5 Comments

Are Developers Faster or Just Busier? It's a question I keep coming back to. AI is clearly increasing output, but more PRs and more commits don't automatically mean a team is moving faster — especially when reviewers and senior engineers are absorbing the cost. Harness is hosting Natalia Sawicka (bol) and James Ford (Compare the Market) at the DevEx Summit on May 13 to talk about how they're actually telling the difference inside their orgs. Worth a spot on the calendar if you lead or work on an engineering team: https://lnkd.in/grVbfhwU

3

Spending hackathon week blurring the lines between 'writing' an integration and 'orchestrating' one. Alongside some incredibly talented colleagues, we’ve been deep in the weeds with Claude Code and multi-agent skills, testing a theory: can a single request turn into a full, assertified reality? It’s early, it’s messy, but we’re starting to see a full 'Artifact Trail' emerge from the noise—telemetry evidence, Knowledge Graph specs, and automated dashboards appearing out of thin air. We aren't just looking at code; we're looking at the potential for self-validating systems. Not sure where this leads yet, but the "what if" is starting to look very real. 🛠️#Grafana #Hackathon #AI #ClaudeCode #MultiAgent #Observability

52

1 Comment

I don't know whether to laugh or cry, but I cannot believe that in 2026 anyone in software development would have the temerity to "...be against unit tests." For the second time in a little over a year I ran into someone with oodles of experience who insisted that unit tests are not truly a "best practice". These were not junior developers. These were folks with multiple decades of experience. I'm close to swearing. For crying out loud, automated unit tests were the first thing that hooked me on AI coding tools. AIs usually write solid tests quickly, and the good tools will learn your style... which I guess assumes that you've written at least a couple it can learn from. It is all I can do to keep from writing this guy's manager or publicly outing him. Sorry, but I just had to vent. Carry on.

146

136 Comments

Some thoughts on job-hunting: There's been quite a bit of anxiety in tech the last few years around AI, job stability, macroeconomic uncertainty... a lot of people with "open to work" on their profiles. To this point I have been lucky and only left jobs/companies on my own terms. So, I can empathize with those looking for work with more uncertainty, but I'm sure I don't fully understand the stresses involved. That said, there are some things I would suggest people keep in mind when they are going through tech interviews: 1. Understand that most people and most companies are very bad at conducting interviews. It is a common point that tech interviews have little to do with actual tech jobs. Almost without exception, no matter who you or how good you are, you will not get an offer from every company you interview with. I have seen some incredible candidates not get offers and get scrutinized in very weird ways. 2. You don't want to get an offer from every company you interview with. You will not be a fit for every company and every company will not be a fit for you. Understanding this will help you take more control over the process and get a better hit rate. 3. Spend less time worrying about what they are looking for and spend more time thinking about what you want to present. I think this is the most important point. I feel this. I see other people struggle with this. You try to figure out what it is the interviewer wants from you and you try to bend to it and you lose yourself. It is important to understand what they want, but it is more important to understand what you want to present. How are you marketing yourself? 4. When they ask "tell us about a conflict you had at work" don't tell them you don't get into conflict. Being an engineer is about solving problems. They are hiring you for your expertise and opinions. If you have opinions you will run into other people with different opinions. They are not asking if you are a jerk. They are asking how you deal with disagreements. If you haven't had any disagreements... go find some. 5. Don't be scared to reframe problems (in some situations). Some of my best interviews have been ones where I didn't know how to do exactly what they were asking. I communicated this and explained how I could get something working if I changed some of the parameters (or they could watch me struggle with something util time was up). I've also had situations where I knew the solution, but could do something more real-world optimal if I changed the parameters a bit. Sometimes I'm told to stay in the lines, but sometimes they're open to it and those have been some of my best interviews. It gives an opportunity to show more of my perspective and experience. 6. Understand that interviewing is a skill you need to maintain in your career. It is different than your day job and takes practice in its own right. You should have a repeatable system/framework for interviews that keeps you focused on what you need to get done.

13

We're wrapping up our migration to Metro internally for Square Android. I summarized some of our challenges and steps we took in a blog post. If you use Dagger 2 for a Kotlin project, then you should migrate to Metro sooner rather than later. https://lnkd.in/g8_iYAKB

58

4 Comments

Developing Luminids is definitely a lot of fun, but for now until early access, it doesn't pay the bills. I'm trying to find work, permanent or contract, and so far I think I'm 150 applications in. This includes hand written, easy apply, tailored CVs, long-drawn out processes on websites (where they ask me my sexual orientation for some reason) across a half dozen websites. Responses so far? ZERO. ABSOLUTELY ZERO. How is this possible? 3 years ago I applied for 2 jobs and got two offers. I thought at my age, the large companies I've worked at and the tech experience I have, I'd be in high demand - instead I'm not even worthy of a phone call apparently. Oof, feels weird honestly. Any help from folks in my network would be greatly appreciated. A share, a like, a post to a recruiter, a kind comment for reach. Here's to making the Luminids dream real and hiring people into the team full time.

16

5 Comments

Ever wonder what separates a senior engineer from a junior in a live coding interview? It's often less about writing the most clever code and more about demonstrating a mature, reliable process. It's about showing how you think. I wrote a new blog, "The Senior Engineer's Playbook for Nailing Live Code Interviews," where I break down the step-by-step framework that great engineers use. From clarifying requirements to testing and iterating. It's a guide to showcasing the problem-solving skills that companies are truly looking for. Read it here: https://lnkd.in/daDwPajm #TechInterviews #SoftwareEngineering #CareerAdvice

26

Are you wanting to learn a new programming language? Well I worked out a way that is very effective. Use the following AI prompt. Replace Clojure with your language of choice. "Teach me Clojure. Let’s work through the basics. Show me something tiny then give me a tiny exercise. After each little thing you show me. Ask me to complete a tiny exercise before we move on to the next thing you’re going to teach me." The AI will then give you some information, and then give you a little exercise. You then paste in the solution for the exercise. The AI will then respond with guidance, and then moves on to the next lesson. E.g. Step 1: Everything is an expression In Clojure, everything is an expression that returns a value. Here’s the tiniest possible thing: Clojure 42 Show more lines That’s a complete Clojure program. It evaluates to the number 42. Now arithmetic. Clojure uses prefix notation (operator first): Clojure (+ 1 2) Show more lines This means: “add 1 and 2” It evaluates to 3. More examples: Clojure (* 3 4) ; 12 (- 10 7) ; 3 (/ 8 2) ; 4 Show more lines General shape you’re seeing: Clojure (operator argument1 argument2 ...) Show more lines Parentheses mean “evaluate this”.

3

What could the AI era look like in a few years? 🤔 With ongoing AI trends, developers increasingly use AI models to get quick solutions to problems. As a result, fewer new solutions, experiences, and discussions are shared on public platforms. A decline in such content could limit the data available for training future AI models. If public knowledge sharing continues to decline, AI models might have less data to learn from, potentially reducing their ability to provide accurate or novel solutions.

7

1 Comment

this makes me think about how I would sell Scala as language. I would say if you want to build something serious, that works and let you sleep at night, confident that I works well, you need types. the strongest/precise the better. this rules out a lot of languages. you need the ability to do whatever you want, so as general purpose as possible and adaptable in every point of your stack. Plus you can decide how much rigorous you want to be: not everything needs to be production ready, but you want to be simple and hacky sometimes and still get correct software. you need FP more then OOP. we all know the pain in having factories, singletons, adapters, patterns over patterns nobody really know how to blend and adapt. you need high level and concise code. you need a mature ecosystem, simple to use especially for intricate tasks like distributed systems and concurrency. you need great tooling and community. that said, if I think about languages that checks these points, scala comes out on top by far as very few others languages have such features. the main problem at this point it's to understand how to invert the downward momentum of scala, as it's less and less adopted/learned. pushing companies away for lack of engineers and just maintain what's there. pushing engineers away for lack of jobs. I'm looking to learn something else myself after seeing the job market by first hand. what can we do about it? pick it as the language to build awesome stuff that can be used by any engineer out there (a little like go or rust are doing, building tools for everyone). I really hope it doesn't die as a language as it's just great

6

I develop video games on the weekends and use AI to help me brainstorm ideas, double-check my thinking, conduct market research, and review code. As a solo developer, one of the most frustrating parts of using AI is not the quality of the answers. It is the lack of memory (context). You can spend hours or weeks working through architecture decisions, debugging edge cases, refining features, and explaining the history of a project. Then the session ends, or you run out of context, and a lot of it is lost. Even if you save to memory before you compact the context, this is only a summary. I have been looking for a way to preserve memory and came across MemPalace (https://lnkd.in/g4vdnRAm). Most AI memory systems try to be selective. They save a few extracted facts and discard the rest. But when you are building something on your own, the value is usually not just in the final answer. It is in the path that got you there. - Why was that approach rejected? - What constraint forced the current design? - What bug showed up last time? - What trade-off did you already think through three weeks ago? As a solo developer, that history matters because there is no team around you to carry that context. If the AI forgets it, you are the one who pays for it in lost time. MemPalace takes a different approach: keeping the actual conversations and making them retrievable. Instead of having AI guess what is important, it stores the raw exchanges and organizes them in a structured way using the memory palace idea of wings, halls, and rooms. In solo development, the conversation is often part of the working record. It is not just chat. It is decision history, technical reasoning, false starts, and lessons learned. Although I am still in the experimental stage with MemPalace, there are a few things I especially like: - It keeps raw conversation history instead of over-compressing everything - It runs locally - It is open source - It does not depend on external APIs

7

1 Comment

Google has just launched a special student offer for all UK-based university students above 18 years old - they can sign up to get the Google Gemini Pro Plan (Google AI Pro) for 12 months, and try new features like Guided Learning, study guides and quizzes. Think of it as a personal AI tutor for every student, helping students understand and learn at their own pace — and not just give an answer. The sign-up period lasts through to 3 November 2025.  More info: gemini.google/students Claim offer: goo.gle/studentsoffer

26

I had a great time chatting with Julie Neumann, Martin Thwaites and Scott Carey for LeadDev yesterday. We discussed how insights into system performance is critical not only for reliability but also for accelerating developer velocity. At the same time, the rise of AI is increasing system complexity and reducing overall comprehension, making observability more essential than ever. From the survey, it’s is also pretty clear that incidents are still the go-to measure for reliability. But should they be? They’re really a lagging indicator when you think about it.

13

I built a cross-platform music sharing app: backend, website, iOS and Android apps... for under £400. Using AI. My background is Linux systems, I wouldn't consider myself a developer. I know Docker, nginx, and Prometheus, but I'd never written a line of TypeScript or React Native. What I learned in the process was that the code was not the key. The real key is how to manage context: a steering document so the AI knows the project, a todo.md so it knows where we left off, and specs before implementation so we agree on what we're building. The result: a Python FastAPI backend with 10 platform integrations, a Next.js website, React Native apps on both stores, CI/CD, monitoring, all running on a single instance. The same project developed by an agency? £30-80k. With AI? Under £500. AI doesn't replace thinking. It replaces typing. You still need to understand your users, design the experience, and own the infrastructure. But the economics of building software just changed fundamentally. Full write-up: https://lnkd.in/eUp9_8ge

38

7 Comments

I spent a weekend building with Lovable. It felt like working with a talented junior dev: fast, eager, and wrong in familiar ways. But still cheaper, better, and faster than hiring one. Where it shines: - prototyping is excellent and fast - genuinely useful for product-minded people, visual preview helps - the 80/20 rule is real, you get to 80% incredibly quick - the code output is solid Where it falls short: - the last 20% is painful, 4 prompts to align 3 buttons - output is decent but boilerplatish, like Tailwind components stitched together - struggles with alignment and spacing (hello junior!) - screenshot-to-code is unreliable, couldn't get even close to a masonry layout after multiple attempts - at some point, I stopped iterating, exporting and fixing in code was faster I burned through 100 credits ($25) in ~4 hours, and got a lot done. 3 years ago this work would've taken at least 1 week, so economics work out. I used Lovable to build the frontend, exported it, then used Claude Code to port it into my existing non-React project. Context matters: I'm a product-minded technical person with high expectations. But for most folks, the output could be just enough. The bar for "good enough" is dropping, and tools like this are democratizing what used to require a developer. I genuinely can't see why you'd pay a frontend developer to create fairly simple layouts from scratch anymore. And these tools will get better and better. It's the end of manual assembly line for frontend implementation, not the end of development. The PSD-to-HTML era is gone, and I'm really happy to see it go :)

18

11 Comments