About
As a Senior Principal Software Engineer, I drive technical excellence, architectural…
Activity
3K followers
-
Serghei Iakovlev posted thisIf you have ever built a RAG pipeline, you have probably faced the same awkward problem: lexical search returns a list of documents with scores like 18.3, your vector index returns another list with cosine similarities between 0 and 1, and somebody on the team asks the obvious question - how do we merge these into one ranking? The naive answer "normalize and add them up" is wrong in interesting ways. The good answer is a sixteen-year-old paper with a one-line formula. That formula is Reciprocal Rank Fusion. It is what Elasticsearch's 'rrf' retriever, OpenSearch's hybrid search, Weaviate, Qdrant, and Azure AI Search all use under the hood. It does not require training. It does not need calibrated scores. It is roughly seven lines of Python. And in the original SIGIR 2009 paper, Cormack, Clarke, and Büttcher showed it beating Condorcet Fuse, CombMNZ, and every individual learning-to-rank method they could throw at it on the LETOR 3 benchmark. For a long time this wasn't on my radar, mostly because the kind of work I was doing didn't force the question. When it finally did, I went looking and found RRF. Wrote up what I learned - link in the first comment. #RAG #SemanticSearch #LLM #MachineLearning
-
Serghei Iakovlev shared thisCoding agents can write code. That's solved. Now try running five of them unattended. In a month of building a Go project with agents handling most of the 242 issues, I hit five failure modes over and over. None involve code generation quality. All involve the feedback infrastructure around the agent. I wrote up the full story in a blog post. Links in the first comment. #OpenSource #CodingAgents #SoftwareDelivery #DevTools
-
Serghei Iakovlev posted thisYour agent writes better TypeScript than Go. Here's why that might not matter. Models generate code based on what they saw during training. JS/TS dominates GitHub and Stack Overflow by volume. Every frontier model scores best on JS/TS. Go has decent but noticeably smaller representation. Agents writing Go get the happy path right, but routinely produce subtle concurrency bugs: goroutine leaks, unbuffered channel deadlocks, error wrapping that compiles fine but loses the sentinel value. But training data is only one dimension. There are others that pull in the opposite direction. Go's verbosity turns out to help. `if err != nil { return err }` is so uniform across codebases that agents reproduce it perfectly. TypeScript has throw, Result types, error-first callbacks, .catch chains, try/catch with async/await. More flexibility, more room for inconsistency. Agents writing Go default to goroutines because concurrency is the path of least resistance. Agents writing Node default to sequential code. The language shapes the architecture the agent produces. I've been thinking about this for a while and wrote up my thoughts. Link in the first comment. #CodingAgents #GoLang #TypeScript #DevTools #SoftwareEngineering
-
Serghei Iakovlev shared thisHow many times have you copied a CI failure log and pasted it back into your coding agent last time? Sortie 1.4.0 closes that loop. CI fails on an agent branch, the orchestrator grabs the logs, feeds them to the next agent turn, agent fixes the build. If it can't fix it after N attempts, it escalates to a human. One config block. No webhook server, no custom scripts. Two agents (Copilot CLI and Claude Code), two trackers (GitHub and Jira), 50 docs pages, 17 releases. Not bad for 17 days. #OpenSource #CodingAgents #DevTools #ClaudeCode #SoftwareDelivery
-
Serghei Iakovlev shared thisI ran 12 coding agents on GitHub Issues simultaneously last month. Three fought over the same file, two got stuck in retry loops, and one burned through $40 before I noticed. The agent wasn't the problem. Everything around it was: workspace isolation, state tracking, failure recovery, cost controls. I was spending more time babysitting agents than writing the code myself. So I built Sortie, an open-source Go orchestrator that watches your issue tracker and dispatches coding agents to tickets. You write a config file, point it at GitHub Issues or Jira, pick your agent, and it handles the rest. Single binary, SQLite persistence. No Postgres, no Redis, no message queue. Kill the process, reboot, come back tomorrow. Every in-flight session and pending retry is still there. Tracker and agent adapters are pluggable: Jira, GitHub Issues, and Claude Code ship today. Swap one, everything else stays the same. 13 releases in 13 days, 41 docs pages, and the GitHub Issues adapter shipped today with v1.1.0. I wrote up the full story in a blog post. Links in the first comment. How are you running coding agents in production? #OpenSource #CodingAgents #GoLang #DevTools
-
Serghei Iakovlev shared thisI’m happy to share that I’ve completed Gen AI Foundational Models for NLP & Language Understanding, the second course in IBM’s “Generative AI Engineering with LLMs” specialization. The course covered core NLP techniques such as one-hot encoding, bag-of-words, embeddings, and neural networks for document classification, along with hands-on labs using PyTorch and torchtext. It was a challenging but rewarding step toward deepening my understanding of how foundational models process and learn from language data. https://lnkd.in/dkhHjuBt #NLP #DataProcessing #Coursera #TextMining #GenAI #DeepLearningCompletion Certificate for Gen AI Foundational Models for NLP & Language UnderstandingCompletion Certificate for Gen AI Foundational Models for NLP & Language Understanding
-
Serghei Iakovlev shared thisExcited to share that I’ve just completed IBM’s “Generative AI and LLMs: Architecture and Data Preparation” course on Coursera. Strengthening my fundamentals is the first milestone on my journey toward more advanced AI work. Looking forward to putting these insights into practice. https://lnkd.in/d5jBiRHR #LLM #NLP #DataProcessing #Coursera #TextMining #GenAI #TransformersCompletion Certificate for Generative AI and LLMs: Architecture and Data PreparationCompletion Certificate for Generative AI and LLMs: Architecture and Data Preparation
-
Serghei Iakovlev shared thisProud to have completed the “Building Agentic RAG with LlamaIndex” course and excited to apply these techniques in real-world AI projects. #LlamaIndex #RAG #LLM #Coursera #DeepLearningAI #GenAISerghei Iakovlev, congratulations on completing Building Agentic RAG with Llamaindex!Serghei Iakovlev, congratulations on completing Building Agentic RAG with Llamaindex!
-
Serghei Iakovlev liked thisSerghei Iakovlev liked this8 вересня буду спікером на AI Tech Rave 2.0 від @Solidgate та DOU Polska у Варшаві. Пройдемося по всіх етапах нашого SDLC і розставимо оцінки: де AI реально змінив темп, а де просто переніс тиск далі. Крім мого виступу, на вас чекають: практичні кейси, досвід та інструменти у доповідях від Solidgate, відверта Ask Us Anything Panel із нашими колегами з Bolt та Netflix, а також афтерпаті — коктейлі, діджей та якісний нетворкінг з неймовірною панорамою на вечірню Варшаву. Якщо ви Software Engineer, Product Manager, QA, Data Analyst,Data Engineer, Engineering Manager, Tech Lead, чи вам просто цікаво, як AI пришвидшує чи сповільнює процес розробки провідних діджитал-продуктів уже сьогодні, — реєструйтеся: https://luma.com/jd0y8nuu Will you vibe with us on AI Tech Rave?
-
Serghei Iakovlev liked thisSerghei Iakovlev liked thisSolidgate held its first tech meetup in Warsaw – AI Tech Rave I can finally wrap up the results of our first tech meetup in Warsaw — it was SO warm and truly high-energy. 🔥 A full room of talented tech professionals, new acquaintances, and conversations that nobody wanted to end. We managed to do a lot in just one evening: talked about why, in the AI era, it's not the idea that wins but the quality — the one thing no prompt can build; broke down how AI is changing the role of a product manager; had a couple of cocktails and exchanged experiences; and even managed to debate whether a product engineer is a product manager who became an engineer, or an engineer who became a product manager. 😄 For me personally, this evening had an extra layer — it was my first ever panel discussion. Honestly, I wasn't sure what to expect, but it went really well, and I have my colleagues and the guests to thank for that. The energy in the room made it easy to open up, go deep, and actually enjoy the conversation rather than just survive it. Warsaw is one of our key hubs in Europe, and the talent and energy in that room showed why. We're here for the long run — building more spaces like this one. P.S. Warsaw, we're just getting started. We've got another major announcement dropping soon. Stay tuned 🔥
-
Serghei Iakovlev liked thisSerghei Iakovlev liked thisSome companies have Claude. We have two. 🙂 Today we're welcoming Claude Alexandre as airSlate's new President. And to clear up the obvious question — no, not the AI. Not the one open in your other tab right now. A real, human Claude, who just happens to share a name with the model half the internet is talking about. Here's what actually matters, though: we didn't bring Claude in because of the AI moment. We brought him in because of how he thinks about customers — the people who use our products every day and the problems they're trying to solve. Claude joined airSlate following more than a decade at Adobe, where he held multiple leadership roles, most recently as Vice President, Digital Media, B2B Product and Campaign Marketing. In that role he led product and campaign marketing across Adobe's Digital Media portfolio — including Creative Cloud, Express, and Document Cloud — helping drive growth for some of the world's most widely used creative and productivity platforms, serving millions of customers worldwide. He also serves as a board member and advisor to several technology companies. As Vadim Yasinovsky, CEO airSlate, put it: "I have been talking to Claude for hours every day...for AI, it is brilliant, transformative and always trying to please me. So, I am thrilled to add another Claude to our team, a human one, who is even more creative and will help transform our entire company. Although I suspect that he will not try to please me as much." And in Claude's own words: “I’ve spent the last 15-20 years building and offering products to companies of all sizes to help them carry out the creative and productivity tasks needed to run their businesses. But however powerful and usable we made them, they’ve always required a lot of work from our SMB customers to achieve their objectives. At airSlate, i have found the dream team and foundation to reinvent how we help those customers with new AI native solutions, and deliver whole outcomes rather than just tooling” Welcome to the team, Claude. Let's head into our next chapter. 💫
-
Serghei Iakovlev liked thisSerghei Iakovlev liked thisLearning in public: Vector Search and Hybrid Search with Reciprocal Rank Fusion ⚛️ Module 2 of the #LLMZoomcamp completed. What it covered: - Turn text into embeddings with a lightweight ONNX model - Minsearch for Vector Search, with chunking for long pages - Compare keyword vs. vector search - Combine them with hybrid search (RRF) 𝗨𝘀𝗲 𝗰𝗮𝘀𝗲: Vector Search isn't always the winner. Ask for an exact token, an error code, a product ID, a rare acronym, and keyword search (BM25) nails the literal match. Hybrid search can get the best out of both worlds. 𝗣𝗿𝗼𝗯𝗹𝗲𝗺: Scores from BM25 and dense vector search live on completely different scales, so you can't just add or average them. 𝗦𝗼𝗹𝘂𝘁𝗶𝗼𝗻: Reciprocal Rank Fusion, a simple idea from Cormack, Clarke & Büttcher (2009). Throw away the raw scores and fuse on rank position instead. Almost no tuning, and it works well enough that Elasticsearch and Databricks use it as a hybrid-search default. More details about RRF in Serghei Iakovlev blog, which I found a great resource: https://lnkd.in/drHrpS-i The LLM Zoomcamp is a free course by Alexey Grigorev & DataTalksClub: https://lnkd.in/dvSA83Ef #learninginpublic #RAG #AI #DataScience
-
Serghei Iakovlev liked thisSerghei Iakovlev liked thisСижу я, никого не трогаю, провожу аудит AI для одной Европейской компании. И тут звонок: «Антон, здравствуйте! У нас менеджер на встрече чашкой с кофе в экран кинул. Надо что-то делать...» Не каждый день поступают такие звонки... Но этот сходу принес несколько тысяч евро. Вам тут может стать интересным сразу несколько вещей: - что такое аудит AI и зачем его проводить - почему консультанту и коучу звонят про разбитый монитор - при чему тут тысячи евро Давайте по-порядку. Нет, мне предложили деньги не за то, чтобы я уволил этого кофе-ниндзя по скайпу и не для того, чтобы я провел ему сеанс управления агрессией. Деньги предложили за то, чтобы прокачать сразу у нескольких сотрудников – одно качество — то самое, без которого внедрение AI рано или поздно превращается в метание посуды. Назову его в этом посте ниже ровно пять раз. Смотрите: Изменения сейчас несутся как стада кенгуру через австралийскую магистраль и выносят даже самые устойчивые команды. Мы с коллегами пару недель назад начали исследовать как AI меняет жизнь сотрудников и бизнеса на деле. Проблем объективно до Луны и обратно. Проникание AI в процессы создает такое давление на людей, что люди хрустят и трескаются как чипсы. И успех определяет качество, которое в отчете Всемирного экономического форума (WEF) Future of Jobs 2025 работодатели поставили на второе место сразу после аналитического мышления и (!) выше лидерства. Это качество еще описывают как связку «устойчивость, гибкость, скорость перестройки», и называется оно «Адаптивность». Вы вот знаете сколько у вас баллов Адаптивности исходя из признанных в мире методов оценки этого качества у сотрудников? То-то же! А это измеряется весьма хорошо и достоверно влияет на успех команды. Очень наглядна компания, которая заказала аудит AI. Три руководителя. Один использует AI, чтобы написать вежливое письмо и как поисковик. Другой прошёл курс по AI, выписал правила, ежедневно тратит пару часов на промптинг. А третий, попивая чаёк, голосом рассказывает ИИ-агенту, чего бы душе хотелось. Дальше агент уже все сам. И смотришь на них – словно эти люди живут в разных десятилетиях. А разница между ними — не в интеллекте, не в возрасте и даже не в знании ChatGPT. Разница — в адаптивности. Словами Дарвина: выживают самые... ну, вы поняли. Вернёмся к разбитой чашке. Первой мыслью было: «Срочно учить их медитации!» (а уж я в медитациях разбираюсь). Потом понял, что ошибся. Эмоциональная усталость и стресс – уже следствие. А я не люблю только симптомы убирать – меня ценят за то, что я причины умею чинть. И вот – вуаля, мы будем системно прокачивать Адаптивность! Измеримо, по шагам, качество, которое имеет поведенческие индикаторы в работе менеджера. Ваши люди выдерживают сейчас темпы изменений? А чего это им стоит? Кому интересны AI-аудиты или внедрение изменений... пишите в личку! А здесь отзовитесь, плиз – кому интересно свою Адаптивность измерить? Всех обнял-приподнял. AI Change Management Consultant | Executive Coach | HRD Expert Антон К
-
Serghei Iakovlev reacted on thisSerghei Iakovlev reacted on thisWhat if the Polish government built GitHub
-
Serghei Iakovlev liked thisSerghei Iakovlev liked thisThe rise of the micromanagers. For years, the common management thesis was : hire a team of talented individuals, set stretch goals for them and then "let them figure it out". Now the same managers using the same philosophy are applying the same kind of thinking to their AI "employees". This is like telling an autopilot to "find its way" from New York to Perth without giving it a map or weather information or telling it that the most obvious way (through the earth) is "outside the scope". For those that haven't had the pleasure, watch "ChatGPT playing chess" by GothamChess (its worth 5-10 minutes and its hillarious). The problem is not that this doesn't work (it doesn't), its that the "capital will solve the problem" works with humans but not with compute. Humans do not have exponential supply curves. It makes no difference if you pay a human $1 or $1B dollars, they will still need sleep, and won't make hundreds of clones of themselves to try to "mean revert" to an answer. This is why we're hitting capacity walls, and why token costs are rising when they should be falling. Capital is not the solution to getting the most out of your AI deployments. The best AI deployments that are production grade are done by micro-managers who train first, build the map, validate intermediate waypoints, QA checkpoints, and only as a final step set "ambitious goals". The next few years will see the rise of the AI-micro-manager class. If you are one of those micro-managers, DM me. If you are one that recently found yourself at odds with management "best practices", don't look "up" at your boss, look "down" at your new AI team who will make you a super-hero. If you're a company that wants to make AI "work" for you : find a micro-manager who everyone hates inside the organization and promote them to be the AI team-lead, you won't be sorry.
-
Serghei Iakovlev liked thisSerghei Iakovlev liked thisWe’re hiring across airSlate. Introducing the first edition of airSlate Open Roles Highlights - where we spotlight priority roles across our teams. In this edition: Engineering, Marketing, Data, and People. If you’re ready for your next move or know someone who might be - take a look below 👇 Featured roles: • Automation QE Lead (SignNow) • Email Lifecycle Team Manager (Marketing) • Senior General QA Engineer (altaFlow) • Senior Analytics Engineer (Marketing Engine) • Senior Data Engineer (Cloud & Security Engineering) • Global People Operations Manager (People Global) Curious to learn more or apply? 👉 Check the comments for direct links Know someone who’d be a great fit? Tag them or share this post 💙 #airSlate #WeAreHiring #OpenRoles #TechJobs #JobsDigest
-
Serghei Iakovlev liked thisSerghei Iakovlev liked thisНарешті закрив цей гештальт… 10 сертифікацій/курсів від Anthropic. — AI Fluency (Framework + Capabilities) — Claude 101 та Claude Code 101 — Claude Code In Action — Introduction to Agent Skills — Intro to MCP + Advanced MCP Topics — Introduction to Subagents — Teaching the AI Fluency Framework Чому пройшов їх, все дуже просто, хочу повністю перейти на Claude замість GPT | Gemini на пів року, але буде видно, бо токени горять - страшне 🫠 #Anthropic #Claude #AIcertification #ClaudeCode #MCP #AIagents
Experience
Licenses & Certifications
Publications
-
Phalcon Cookbook
Packt Publishing
Phalcon Cookbook is a 386-page book that teaches new and intermediate Phalcon users how to use and master Phalcon, Phalcon ORM and Phalcon DevTools.
Other authors
Courses
-
Basic to Advanced: Retreival-Augmented Generation (RAG)
-
-
Building Agentic RAG with Llamaindex
-
-
C++ White Belt
-
-
C++ Yellow Belt
-
-
Functional programming in Haskell
-
-
Fundamentals of AI Agents Using RAG and LangChain
-
-
Gen AI Foundational Models for NLP & Language Understanding
-
-
Generative AI and LLMs: Architecture and Data Preparation
-
-
Kotlin Developer
-
-
Python Core
-
Projects
-
Sortie
- Present
Sortie turns issue tracker tickets into autonomous coding agent sessions. Engineers manage work at the ticket level. Agents handle implementation. Single binary, zero dependencies, SQLite persistence.
-
setup-specmatic
setup-specmatic is an action to set up a Specmatic environment for use in actions by optionally downloading and caching a version of specmatic by version and adding to PATH.
-
Ed Theme
Ed is a Hugo theme designed for textual editors based on minimal computing principles, and focused on legibility, durability, ease and flexibility.
-
django-environ
A Python package that allows you to use Twelve-factor methodology to configure your Django application with environment variables.
I'm a maintainer, Daniele is the original author.Other creatorsSee project -
esup
A GNU Emacs Start Up Profiler.
My contributions included stability fixes, unit tests, CI/CD, translations, maintaining project, PR reviews and more.Other creatorsSee project -
jaeger-client-php
A client-side library that can be used to instrument PHP apps for distributed trace collection, and to send those traces to Jaeger.
I'm a maintainer.Other creatorsSee project -
Zephir Language
-
Zephir - is a high level programming language that eases the creation and maintainability of extensions for PHP. Zephir extensions are exported to C code that can be compiled and optimized by major C compilers such as gcc/clang/vc++. Functionality is exposed to the PHP language.
My contributions included:
★ Contributed to performance optimization and improved compilation speed, ensuring faster and more efficient builds.
★ Diagnosed and resolved critical bugs in the compiler and…Zephir - is a high level programming language that eases the creation and maintainability of extensions for PHP. Zephir extensions are exported to C code that can be compiled and optimized by major C compilers such as gcc/clang/vc++. Functionality is exposed to the PHP language.
My contributions included:
★ Contributed to performance optimization and improved compilation speed, ensuring faster and more efficient builds.
★ Diagnosed and resolved critical bugs in the compiler and kernel, enhancing system stability and reliability.
★ Designed and implemented new language features to extend functionality and usability for developers.
★ Developed and maintained unit tests to ensure robust, high-quality code and prevent regressions.
★ Oversaw project maintenance, including code review, repository management, and release coordination.
★ Conducted in-depth code reviews to uphold coding standards and improve overall project quality.
★ Managed project releases, including versioning, documentation updates, and feature rollouts.
★ Acted as a core developer and co-owner, collaborating with team members to define and execute project goals.Other creatorsSee project -
Zephir IDEA plugin
-
IntelliJ plugin for editing Zephir code. Provides syntax definition, autocompletion and syntax check support.
Other creatorsSee project -
Phalcon Framework
-
An open source web framework delivered as a C extension for the PHP language providing high performance and lower resource consumption.
My contributions included:
★ Contributed as a Core Team Member, Maintainer, and Co-Author to the development of Phalcon, a full-featured PHP framework delivered as a C extension.
★ Optimized framework performance, improving execution speed and efficiency for high-demand applications.
★ Implemented security and stability fixes to ensure robust…An open source web framework delivered as a C extension for the PHP language providing high performance and lower resource consumption.
My contributions included:
★ Contributed as a Core Team Member, Maintainer, and Co-Author to the development of Phalcon, a full-featured PHP framework delivered as a C extension.
★ Optimized framework performance, improving execution speed and efficiency for high-demand applications.
★ Implemented security and stability fixes to ensure robust, secure, and reliable framework releases.
★ Developed and maintained unit tests to uphold code quality and prevent regressions.
★ Packaged and maintained framework distributions for Ubuntu and Windows, ensuring cross-platform compatibility.
★ Designed and implemented CI/CD pipelines to streamline testing, builds, and deployments.
★ Provided project translations to expand the framework's global accessibility and user adoption.
★ Maintained and managed the project lifecycle, ensuring consistent updates, stability, and feature rollouts.
★ Served as the primary maintainer of the Zephir language, enabling efficient PHP-to-C compilation for Phalcon.Other creatorsSee project -
Phalcon Cookbook
-
I wrote a 386-page book co-authored with David Schissler on Phalcon. It's called "Phalcon Cookbook" and it has, since its release, sold thousands of copies.
Other creators
Languages
-
Russian
Native or bilingual proficiency
-
English
Professional working proficiency
-
Ukrainian
Limited working proficiency
-
Polish
Limited working proficiency
Organizations
-
Codeception Testing Framework
Core Team Member
- PresentA full-stack testing PHP framework.
-
Phalcon
Core Team Member
-An open source web framework delivered as a C extension for the PHP language providing high performance and lower resource consumption.
Recommendations received
11 people have recommended Serghei
Join now to viewExplore 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