标签：#softwareengineering

Inspired by Nir Eyal's "beliefs are tools" framework Beliefs are tools, not truths. Tech stacks are too. Pick the ones that work for you. Most "tech debt" is actually "belief debt". We hold onto frameworks, patterns, and processes long after they stop serving the product. To build great software, we need to introduce a core rule: If a tech belief or "best practice" doesn’t solve a real problem for you right now, it must be treated as false. Here is how to audit your tech beliefs using 5 filters. 1. ARE THEY USEFUL? The real question isn’t "Is this the best tech?" It’s "Does this serve the user?" Tools are tools. Keep the ones that ship. Bad belief (Treat as False): "We need Kubernetes because it’s the industry standard." Useful belief (True for Now): "A $5 VPS serves 10k users. We’ll use K8s when we have a scaling problem, not a resume problem." If your architecture choice doesn’t make the core loop faster, cheaper, or simpler for users, it’s not serving you. Delete it. 2. ARE THEY TESTED? A useful stack holds up when the world pushes back. Pay attention to production, not the trending blog posts. Bad belief (Treat as False): "Microservices are inherently more scalable"—said before you even have 2 concurrent users. Tested belief (True for Now): "Our monolith handles 50 req/s perfectly. We’ll split services only when latency exceeds 300ms in prod." Load test it. Dogfood it. If it only works in a conference slide deck, it’s a story, not a tool. 3. ARE THEY OPEN? A tech choice you can’t change has stopped being a tool and has become a cage. Hold opinions firmly, but hold implementations loosely. Bad belief (Treat as False): "We’re a React shop forever." Open belief (True for Now): "React serves us today. If HTMX lets us ship this feature in 2 days instead of 2 weeks, we’ll use HTMX." In a famous study on hope, Curt Richter’s rats swam for 60 hours when they believed rescue was coming. Your team will grind for years on a legacy stack if they believe it can actually be r

A practical reflection on why coding agents lose the thread between sessions, and why the repository itself is the right place to preserve it. I used to think the problem was memory. That was the obvious diagnosis. Every new coding-agent session started with the same ritual. Open the repository. Read the README. Inspect the project structure. Search for the files that looked important. Reconstruct the task. Guess which commands mattered. Ask again what had already been tried. Then do the actual work. Sometimes. Because a lot of the work was not work. It was orientation. A coding agent can have a large context window and still lose the operational thread. It can have chat history and still fail to know what happened in the last run. It can retrieve semantically similar notes from a vector store and still miss the one fact that mattered: this command already failed. the previous session stopped here. this file looked relevant, but it was a dead end. the validation was not actually run. One day I stopped thinking about the problem as "agent memory". That word was too broad. Too attractive. Too dangerous. Because once you say memory, the temptation is to build a bigger one. A bigger context window. A bigger note store. A bigger vector database. A bigger archive of everything the agent has ever seen, said, touched, generated, or vaguely implied. That sounds powerful, but it is also how you build a very expensive junk drawer. Context is not continuity Context is what the agent has available now. Continuity is what lets the next execution continue from what actually happened before. Those are not the same thing. Long context helps while a session is alive. It gives the model more text to work with. More files. More prior messages. More implementation details. More room. Although it is really useful it does not automatically produce continuity. When the session ends, gets compacted, moves to another tool, switches from one coding agent to another, or simply starts tomorrow

imagine you are working in a large codebase. you need to fetch different kinds of data and transforming them, grouping them or sorting them. lets take a closer look at Sorting and Heaps in golang Specifically. we already familiar with heaps and its important interface. the Heap.Interface. a very performant and impressive implementation of heaps and its sorting functionality. type Interface interface { sort . Interface Push ( x any ) // add x as element Len() Pop () any // remove and return element Len() - 1. } as a programming language, it couldnt be done better than what it is today. but most of the time we might not need to implement all the interface items. dont get me wrong the functionalities should exist but mostly all that matters for us is that how the sorting will be done. ZenQL's Implementation In the latest version take advantage of sorting and heaps functionality. in a fast and agile way! result := From ( personList ) . Where ( func ( person Person ) bool { return person . Active == true }) . CollectSorted ( func ( person Person , person2 Person ) bool { return person . Identifier < person2 . Identifier }, true ) In the code snippet above we perform a sort on our collections using the thor engine very easily. we just express our desire about how the sorting needs to be done and wether its ascending or descending. and other functionalities are implemented as below: type Sortable [ T any ] struct { Items [] T less func ( a , b T ) bool desc bool } func ( h Sortable [ T ]) Len () int { return len ( h . Items ) } func ( h Sortable [ T ]) Swap ( i , j int ) { h . Items [ i ], h . Items [ j ] = h . Items [ j ], h . Items [ i ] } func ( h * Sortable [ T ]) Push ( x any ) { h . Items = append ( h . Items , x . ( T )) } func ( h * Sortable [ T ]) Pop () any { old := h . Items n := len ( old ) item := old [ n - 1 ] h . Items = old [ : n - 1 ] return item } be faster and more agile with the Golang ZenQL. Click To Visit ZenQLRepository

Last week, Toronto hosted Tech Week. A city-wide celebration filled with events and workshops...

The hidden cost of AI isn't generating code. It's understanding your codebase. For a long time, I assumed AI coding tools became expensive because they generated a lot of code. These tools can produce components, tests, SQL queries, documentation, and sometimes entire features on demand. If costs were climbing, the output volume must be the reason. The more I used these tools, the more I realized I was measuring the wrong thing. The expensive part isn't writing code. The expensive part is understanding what code should be written — and that work is mostly invisible. That realization changed how I think about AI-assisted development entirely. Two Prompts, Two Very Different Problems Consider these two requests: "Create a utility function that formats dates" and "Review this feature and suggest improvements." At first glance, both look ordinary. Both might even produce short answers. But they require completely different levels of understanding. The first is narrow and well-defined. The AI needs very little information before it can produce a useful answer. The second is open-ended. Before suggesting a single improvement, the AI may need to read multiple files, understand dependencies, follow existing patterns, compare implementations, and build a mental model of why the feature exists at all. The output might still be small. The work required to reach it is not. Why Agent Workflows Feel Different From Autocomplete This became much clearer when I started using AI agents. Traditional autocomplete is predictive — you type, the AI guesses what comes next. It's fast, cheap, and deliberately context-light. Agents behave differently. When you ask one to improve a feature or review a workflow, it doesn't immediately start generating code. It starts reading. It follows imports, finds related files, and tries to understand the system before touching it. That is exactly what makes agent workflows feel slower and more resource-intensive than autocomplete: they are spending effor

CAP Theorem Explained: Choosing Between Consistency, Availability, and Partition Tolerance in Databases Imagine you're trying to book a flight online, and just as you're about to pay, the website crashes. When you try to book again, you find that the flight is now sold out, even though the website initially showed available seats. This frustrating experience is a classic example of a database trade-off between consistency, availability, and partition tolerance. The CAP theorem, first introduced by Eric Brewer in 2000, states that it's impossible for a distributed data store to simultaneously guarantee more than two out of these three principles. In this post, we'll delve into the world of CAP theorem, exploring its fundamentals, real-world database examples, and design implications. Introduction to CAP Theorem Understanding the Basics of CAP Theorem The CAP theorem is based on three primary principles: Consistency : Every read operation will see the most recent write or an error. Availability : Every request receives a response, without guarantee that it contains the most recent version of the information. Partition Tolerance : The system continues to function and make progress even when network partitions (i.e., splits or failures) occur. Importance of CAP Theorem in Distributed Systems In distributed systems, where data is spread across multiple nodes, the CAP theorem plays a crucial role in understanding the trade-offs between these principles. By grasping the CAP theorem, developers can design more resilient and scalable databases that meet the specific needs of their applications. Brief Overview of the Blog Post This post will explore the CAP theorem in depth, using real-world database examples to illustrate the trade-offs between consistency, availability, and partition tolerance. We'll discuss the fundamentals of CAP theorem, examine CA, CP, and AP systems, and provide guidance on designing for each combination. By the end of this post, you'll have a solid un

Most API teams treat the contract as an afterthought: write code, generate a spec, then watch the two drift apart. Git-native API design reverses that flow. You treat the API contract as source code, version it in Git, and review every change the same way you review application logic. Try Apidog today This guide focuses on implementation discipline, not a single tool. You’ll design contracts in branches, review them in pull requests, and turn a committed spec into mocks, tests, and docs. The goal is simple: your Git history should also be your API history. If you already know what Spec-First tooling looks like and want the product walkthrough, read the companion piece on the git-native API workflow . This article stays focused on practice. What “git-native” means for API work Git-native means your API definition lives in your repository as a plain text file. Not in a proprietary cloud database. Not behind a vendor login. A .yaml or .json file sits next to your code and is tracked by the same version control system your team already uses. In many cloud-locked API design tools, the contract lives in the vendor’s backend. You edit through a web UI, and your repository only contains an export. That export can become stale, and your Git history no longer explains how the API evolved. The git-native model inverts that relationship: The file in main is the contract. Any GUI is a view onto that file. Branches, commits, pull requests, blame, and rollback all apply to your API surface. Mocks, docs, tests, and generated clients derive from the committed spec. A git-native setup has three core properties: The spec is a text file in the repo. Changes flow through normal Git operations: branch, commit, PR, merge. Downstream artifacts derive from the committed file, not from a separate database. Why design and develop APIs in Git You already trust Git with your code. Your API contract deserves the same treatment. 1. History When someone asks, “When did we add the cursor pagination

A security reviewer finds a critical issue a day or two before the release of an application. While it's an important issue, it sets the team back weeks, frustrating their product management partners and customers. The review came at the most expensive time in the process. There are many examples of how work items move through different processes to deliver software in large companies. While GenAI has allowed us to rapidly create code, it also moved and exposed the bottlenecks in our processes. It has also caused us to re-examine where it is most effective to make certain decisions. This is the challenge, and a deliberate blend of automated, programmatic, and human judgment is well suited to help you solve it. We can borrow from the well-trodden path of value stream mapping here. It is useful for spotting bottlenecks and waste in a given process, but it's also valuable to ask the deeper question of who or what should own each step. Each option earns its place differently. Is there an earlier step that may reduce costs with an agent where it was previously limited by human availability? Or is the stronger determinism of a programmatic step more important for a critical piece of the flow? Some decisions should stay with human judgment, where confidence without context is a liability. The opportunity for security teams and other stakeholders is to scale their impact across these options rather than scaling headcount. Workflow-as-code is not a new idea. There are a number of existing engines where the workflow definition is its own entity, separate from the work itself. GitHub Actions defines pipelines in version-controlled files, while the execution happens on separate runners. Airflow and Temporal follow a similar pattern for data and application workflows. Because the definition lives on its own, a team can change how a given step runs without rebuilding the whole flow. That separation is what makes it practical to adjust who or what owns each step over time. Rather

You know the moment. You push the branch, open the PR, and immediately see it — the undefined return on the refund path, the token logged to the console, the TODO that was supposed to be temporary six weeks ago. The reviewer catches it four hours later and you reply "good catch, fixing now" as if someone else wrote that line. The first reviewer on most pull requests should have been the author. Half the comments you will receive — the missing null check, the untested error branch, the duplicate logic that could be extracted, the import that now goes nowhere — are things you would have caught with one more careful read-through. You skip that read because you have been in the code for two days and your brain completes the sentences for you. You see what you meant to write, not what is on the page. This post is about closing that gap with a structured AI-assisted self-review before the PR opens. Not to skip the human reviewer — to walk into the review with the obvious problems already gone, the test gaps already filled, and the PR description already written. So the reviewer's attention can land on what actually needs a second pair of eyes. The tool is branchdiff : a local browser app that runs your diff on localhost , stores everything in ~/.branchdiff/ , and keeps the AI surface controlled through an explicit branchdiff agent command API. Nothing leaves your machine until you decide to push it. Why "before the PR" is the right moment If you review after opening the PR, every AI fix becomes noise: a force-push, a re-read for your reviewer, another commit in the audit trail. If a teammate is already mid-review when you discover the bug, you look careless. The patch that should have been in the original push becomes a distraction for everyone downstream. If you review before opening the PR, the AI's output is a private workspace. You act on what matters, commit the fixes into your own history (often as fixup! commits you squash before pushing), and the PR that goes up i

The loop I didn't notice closing Seven weeks ago I started using AI for work. Two weeks after that, I published an article. Seven weeks after that — today — the article is one of sixteen, and they are all in a memory file that the AI reads at the start of every new conversation. I didn't notice the loop until I named it. This is a note about that loop, what it is, what it isn't, and why I keep publishing even though the loop doesn't strictly need me to. The shape It runs like this: I decide what to do. I work it out with the AI — usually in dialogue, sometimes by pasting raw code or data. The dialogue becomes a record. Sometimes a memory entry. Sometimes a published article. The record becomes context for the next conversation, which informs the next decision. It didn't look this clean while it was happening. The numbering is hindsight. From inside, the steps overlap. The first step is the one I keep. Direction is mine: what to build, what to write, what to negotiate. The history that shapes those decisions — twenty-four years of solo work, my company, my family, my health — is also mine. The AI is not setting direction. The second step is where most of the leverage is. I describe what I want to do as completely as I can, sometimes by handing over source code. Then I ask: does this look right? Is there a path I'm missing? Where would this break? I'm opening drawers — possibilities I half-saw in my own head — and checking which ones open cleanly. When one opens cleanly, that is the GO signal. Not "will this succeed" but "this is doable, so do it." The third step happens almost without effort. The conversation already exists as text. Some of it becomes a memory entry I add deliberately. Some of it becomes raw material for an article. The article writes itself partly because I have already explained the thing to the AI. The fourth step is the one that took longest to arrive — and the one I want to be most careful about describing. Three phases, not one The loop didn't

Over the past few months, I’ve increasingly noticed something through my network: more people from non-technical backgrounds are building software as AI tooling improves. Designers are prototyping product ideas. Product managers are testing workflows. Founders are building MVPs. Operators are creating internal tools. People who would not have called themselves “technical” a year ago are now using AI to make ideas tangible. I think this is genuinely exciting. It has never been easier to create. I even attended a hackathon where participants only had 20 minutes to build a demoable product! This raises the question: When AI makes building easier, how do we make sure understanding does not disappear? I recently published Thinking in the Age of AI , a guide for software engineers (you can check out my previous post here ). That guide focused on individual reflection for engineers: how to keep developing technical intuition, reasoning, and judgment while using AI tools. But the landscape has changed quickly. AI-assisted building is no longer only an engineering workflow. It is becoming a builder workflow accessible to all. And by builders, I mean anyone using AI to turn ideas into software-like artifacts: vibe coders designers product managers founders operators marketers students non-engineering team members So I wanted to create a new version of the system for this wider builder audience. Thinking in the Age of AI: Builder Edition The opportunity is real I do not think we should dismiss this shift. I have spoken with people from all kinds of backgrounds who are actively building now. People who previously had to wait for engineering time can now create something concrete. That changes the conversation. Instead of describing an abstract idea, you can show a flow. Instead of writing a long product spec, you can prototype the interaction. Instead of asking “would this work?”, you can test a rough version. That is powerful. But there is a trap. A prototype can look much mor

How a Product Sync Automation Project Transformed Customer Onboarding When people think about impactful engineering work, they often imagine distributed systems, high-scale infrastructure, or complex algorithms. One of the most impactful projects I worked on wasn't any of those. It was solving a seemingly simple problem: Keeping product data in sync across multiple retail systems. Years later, our CEO still remembers how much smoother customer onboarding became after this project. The Context: What is Commerce Connect? At Casa Retail AI, we have an internal platform called Commerce Connect (CC) . Commerce Connect acts as the central Product Information Management (PIM) system and serves as the source of truth for product information. Under the hood, it is built on top of a customized version of the open-source e-commerce platform Spree Commerce , extended with multi-vendor and multi-tenant capabilities. Its primary responsibility is simple: Collect product information from multiple retail ecosystems and distribute it to every Casa product that needs it. Once product data enters Commerce Connect, it is synchronized to multiple downstream systems. Why Product Data Matters Many applications inside Casa depend on product information. Product Consumers Once product data enters Commerce Connect, it is distributed to multiple systems across the Casa ecosystem. Customer-Facing Applications Several products rely on product information to provide context and improve customer experience: Lead management applications use product information during customer interactions. Ticket management systems link customer issues to specific products. Digital receipts display product names, images, and related details. Analytics & Reporting Product data powers business dashboards and reports, helping retailers answer questions such as: Which categories perform best? Which products attract the most attention? Which products generate the most complaints? It is also used for filtering and segme

Agentic AI coding tools are transforming how we build software. But they share a fundamental constraint: context windows are finite, and as chat sessions grow, AI performance degrades, a phenomenon Anthropic calls context rot . The model loses its grip on early instructions, leading to a frustrating "fix-it loop" where the agent fixes one thing but breaks another. Most of us prompt an agent, let it write code, review it, and repeat. This works beautifully for prototypes. But when you need to build a stable, full-featured product with hundreds of mission-critical acceptance criteria (AC), "vibe-coding" breaks down. The reality is that you get better behavior from agents the same way you get it from humans, by explicitly capturing what good and bad look like, and checking against it . Coming from a systems engineering background in regulated industries, I knew we needed to stop treating agents like conversational chat buddies and start treating them like engineering assets. That's why I built DevCortex : a purpose-built structured intelligence layer that brings systems engineering discipline to agentic workflows. What is DevCortex? DevCortex is an agentic development platform built on one core idea: AI agents work best when they have structured, queryable access to a database of requirements they can interrogate on demand, not a wall of text in a prompt. It sits between the human specification and AI execution using three components: 1. An Agentic-V Model Database: A structured hierarchy mapping your high-level vision (ConOps) to system specs (Specs), individual requirements (Reqs), linked defects (Issues), and an auto-generated Traceability Matrix. 2. An MCP Server: Delivers just-in-time, high-signal context to tools like Claude Code or Open Code. Instead of dumping requirements upfront, the agent queries exactly what it needs, when it needs it. 3. Human Control Planes (Web UI & CLI): A multi-user Web UI with real-time WebSocket feeds to watch your agent work, plus a

Most software projects focus on producing outputs. Most AI projects focus on producing predictions. But real organizations don't operate on outputs or predictions alone. They operate on decisions. A decision has consequences. A decision creates risk. A decision consumes resources. A decision changes the future state of a system. Over the last few months, I've been studying and building systems around a simple question: How can we make decisions more explainable, auditable, and repeatable? This led me toward concepts such as: event-driven architectures decision logging risk evaluation pipelines audit trails feedback loops operational intelligence systems Instead of asking: "Can we predict what will happen?" I'm becoming more interested in asking: "Can we explain why a decision was made?" and "Can we reproduce that decision six months later?" Current areas I'm exploring: Financial decision systems Risk infrastructure Event-driven architectures Blockchain compliance workflows Operational intelligence platforms One of the projects I'm currently building is an Event-Driven Decision Logging System (EDDL), designed to explore how organizations can record, audit, and replay critical decisions over time. Still learning. Still building. Still refining my understanding of how complex systems operate under uncertainty. Looking forward to sharing the journey here. systemsdesign #architecture #backend #fintech #softwareengineering #eventdriven #riskmanagement

I understand why so many people are questioning software engineering right now. Every week there’s another headline saying AI will replace developers. Junior engineers are worried there won’t be jobs. Senior engineers are wondering how long their experience will stay valuable. And honestly, if you spend enough time on tech Twitter or LinkedIn, it can start feeling like the industry is collapsing in real time. But after using AI heavily in my day-to-day work as a software engineer, I’ve started seeing things differently. AI didn’t make me feel less useful. It made me feel more capable. Before AI became part of my workflow, a lot of engineering time disappeared into things that were mentally draining but necessary: repetitive refactoring debugging small issues writing boilerplate digging through documentation trying to remember syntax cleaning up legacy code writing SQL queries optimizing simple functions translating vague tickets into technical tasks None of these tasks were impossible. They were just time-consuming. Now, a lot of that friction is reduced dramatically. One of the biggest changes I noticed was backlog cleanup. Tasks that used to sit untouched because nobody wanted to deal with them suddenly became manageable. Not because AI magically solved everything. But because it helped reduce the “mental startup cost” of difficult tasks. Sometimes all you need is: a starting point a refactored example help understanding unfamiliar code a faster debugging path quick documentation summaries That momentum matters more than people realize. A task that feels overwhelming at 9AM suddenly becomes achievable when AI helps break it down. I also noticed we started delivering faster as a team. Not in a “replace developers with AI” kind of way. More in a: less context switching faster research quicker prototyping fewer hours stuck on repetitive problems better ticket breakdowns improved communication kind of way. The interesting part is that AI didn’t just help with coding.

Something broke in the AI pricing market between January and May 2026. A year ago, "frontier model" meant "expensive model." Claude Opus was $15/$75 per million tokens. GPT-4 was $5/$15. If you wanted the best coding performance, you paid the best price. The correlation between quality and cost was loose, but it existed. That correlation is gone. The Numbers That Changed Everything Here's SWE-bench Verified — the benchmark that tests AI models against real GitHub issues from projects like Django, Flask, and scikit-learn — plotted against output price per million tokens: Model SWE-bench Output $/1M Score/Dollar ───────────────────────────────────────────────────────────────── Claude Opus 4.7 87.6% $25.00 3.5 Claude Opus 4.6 80.8% $25.00 3.2 Gemini 3.1 Pro 80.6% $15.00 5.4 GPT-5.2 80.0% $10.00 8.0 DeepSeek V4 Pro (Max) 80.6% $3.48 23.2 Kimi K2.6 80.2% $4.00 20.1 Qwen3.6 Plus 78.8% $3.00 26.3 MiniMax M2.5 80.2% $1.20 66.8 DeepSeek V4 Flash (Max) 79.0% $0.28 282.1 Read that last line again. DeepSeek V4 Flash scores 79% on SWE-bench at $0.28 per million output tokens. Claude Opus 4.7 scores 87.6% at $25.00. The performance gap is 8.6 percentage points. The price gap is 89x . For a team running 100 million tokens per month, that's the difference between $28/month and $2,500/month. For a 9-point improvement in code completion accuracy. It's Not Just One Outlier This isn't a DeepSeek anomaly. Look at the cluster of models scoring 78-80% on SWE-bench: DeepSeek V4 Pro : $3.48/1M output — open source, 1M context Kimi K2.6 : $4.00/1M output — open source, 256K context MiniMax M2.5 : $1.20/1M output — open source, 200K context Qwen3.6 Plus : $3.00/1M output — open source, 1M context MiMo-V2-Pro : $3.00/1M output — open source, 1M context Five models from five different Chinese labs, all scoring within 2 points of GPT-5.2 ($10.00/1M) and Gemini 3.1 Pro ($15.00/1M), all at 1/3 to 1/10 the price. And they're all open source. What Happened Three things converged: 1. Mixture-of-Exper

Over the past few years, my workflow as an engineer has changed a lot. I went from the occasional...