Devsu

Devsu mentioned by Gartner in a Magic Quadrant. 👏 In short: We offer AI to accelerate software modernization for the enterprise. It's producing astonishing results and I know this is just the beginning. Congrats to Felipe Cornejo and the Devsu team, specially to our AI Innovation Hub!

Andrej Karpathy just published his 2025 LLM year-in-review, and one concept stopped me cold: "jagged intelligence." It reminded me of a scene in the Netflix K-drama "Extraordinary Attorney Woo," where the brilliant autistic lawyer Woo Young-woo struggles with a revolving door. She'd just delivered a flawless legal analysis—sharp, precise, world-class reasoning. Then she stood confused at this simple rotating entrance, unsure how to navigate something most people do without thinking. That's exactly how LLMs work now. We're not evolving smarter animals. We're creating a fundamentally different kind of intelligence—one optimized for entirely different pressures than human cognition. One that can crush complex math proofs but get tricked by simple jailbreaks. Solve PhD-level problems but stumble on grade-school logic. The smartest polymath and a confused child—simultaneously. Karpathy highlights a few paradigm shifts from 2025: Reinforcement Learning from Verifiable Rewards (RLVR) became the new standard training stage. LLMs learned to "reason" through trial and error against objective rewards, not human templates. The capability jump was obvious. Claude Code showed AI agents living on your computer, not just in cloud containers. With your context, your data, your environment. They got it right by going to your computer first, not to the cloud. Vibe coding crossed the capability threshold. Even Karpathy—who knows how to code properly—now builds custom Rust tokenizers and ephemeral debugging apps through English alone. Code became free, malleable, disposable after single use. But here's what matters most for enterprise: we haven't realized 10% of LLM potential at current capabilities. The technology already exists. The bottleneck isn't AI processing speed—it's human learning speed. At Devsu, we're building this layer Karpathy describes: the context engineering layer. With Velx, we're packaging the guardrails, standards, and compliance requirements enterprises need to use LLMs safely for app modernization. Not generic AI assistants—specialized agents with the right context, trained on your patterns, aligned with your standards. Because brilliant doesn't mean universally capable. It means knowing where intelligence spikes—and where the revolving doors are. The question isn't whether AI transforms software development. The question is: are you building the context engineering layer to make it actually work? Read Karpathy's full year-in-review: https://lnkd.in/eTij23UD What surprised you most about AI capabilities this year? #ArtificialIntelligence #SoftwareDevelopment #ApplicationModernization #EnterpriseAI #Leadership

Everyone optimizes for growth and profit. I optimized for something else. It's been the better strategy. I've watched the startup world operate on a seductive logic: Move fast and break things... but sometimes you end up breaking people. The efficiency-first mindset is everywhere. And I'll be honest—I've felt the pull. The pressure to optimize for output over people is constant. Some seasons I've leaned too far that way. Last Saturday, my wife and I hosted our Devsu team for Thanksgiving at our home in Quito. Watching everyone gather, share, and genuinely connect reminded me how much treating people well matters. I kept coming back to what Jesus actually said about leadership. Not the corporate-sanitized version—the real thing. "Whoever wants to be great must be a servant," "Love your neighbor as yourself," and "Do to others what you would want them to do to you." Here's what I learned: people have perfect radar for authenticity. They can feel whether you care about their growth or just their output. Whether you're building them up or using them up. When you genuinely love your team—when you invest in their growth even when it's inconvenient, when you celebrate their wins even when that means they won't be working with you anymore, when you build systems that serve them and not just the bottom line—something shifts. Retention goes up. People stay because they feel appreciated, not just because of salary. Quality goes up. People give their best when they know you want their good, not just the company's good. And legacy becomes possible. Love builds something that outlasts you. Transactions don't. Caring for people isn't soft leadership. It's the ultimate competitive advantage. Because in a world optimizing for extraction, genuine love stands out like a lighthouse. What do you prefer? Winning first place by breaking everyone around you? Or compete fairly and work with people that feel appreciated and valued? Jesus knew what he was talking about. I'm still on this learning path, and I've made tons of mistakes, but love is the path I want to follow. What about you?

Yesterday I attended leadership talks at AI Engineer Code Summit. My brain is still processing. I'll be posting a series on my learnings — this is a preview with the first 4 talks. Here's what the people building the future of software development are actually doing: 🧠 Evolving Claude APIs for Agents (Anthropic — Katelyn Lesse) There's a 39% performance improvement in LLMs just by managing context window. Not better models. Not more parameters. Just better context. Claude API has very nice tools: thinking budget, memory, context editing, code execution sandboxes, and what they call "agent skills" — curated instruction folders for specific domain knowledge. The takeaway: before you chase the next model upgrade, clean up your context. 🤖 Autonomy Is All You Need (Replit — Michele Catasta) Replit is making code creation reliable for non-technical users. A key insight on building autonomous agents: Agents perform better on narrow scope. And without automated testing, they build "painted doors" — things that look finished but aren't functional. Their solution: agents write Playwright tests instead of manually checking their work. Honest validation + a regression suite you didn't write. Also: 200k token windows are more than enough. The secret isn't bigger context — it's subagents that protect the core agent's focus. 🔧 Your Support Team Should Ship Code (Zapier — Lisa Orr) This one challenged the assumption that support teams cannot deliver code to production. Zapier empowered their support team to deliver code fixes. The agent analyzes tickets, determines if they're auto-fixable, writes the code, and support reviews and creates the PR. Engineering only reviews. Result: Each support team member ships 3-5 fixes per week. The insight: they're closest to the client and the problem. Why not let them fix it? ⚡ 2026: The Year the IDE Died (Gene Kim & Steve Yegge) Same as me, Steve spends a lot of time in Claude Code and enjoys it! Cool! 🙌 The evolution: Traditional Coding → Prompting → Agentic Coding → Orchestrators. Their advice: "Until AI orchestrators arrive, use agentic coding now. It will smooth the transition." "The more you use AI, the more you trust it." "AI will reshape tech orgs 100x larger impact than agile, DevOps, cloud" A. Cockcroft once talked about NoOps. Now we talk about NoDev. Quoting Erik Meijer: "If we are the last generation of developers to write code by hand, let's have fun doing it.". 😁 What This Means for Us At Devsu, we're already applying these principles with Velx, our AI-native application modernization solution. We've been working hard on Context Engineering, crafting context that matters for each step of the process — curated instruction sets that give our agents the context they need without drowning them in irrelevant information. Progressive disclosure, not information overload. The result: up to 70% reduction in app modernization time. Do you want me to continue this series? Comment below!

Yesterday I met Gene Kim. If you've worked in software development over the past decade, you know what The Phoenix Project did for our industry. It fundamentally changed how we think about the software development lifecycle. I've applied countless lessons from that book at Devsu. I had maybe two minutes with him. I told him about the impact his work had on me, thanked him for his contribution to the field. He was signing copies of his newest book, "Vibe Coding," co-written with Steve Yegge. Got both signatures 😁 . Gene told me something interesting: "This is my best work yet." Coming from the author of The Phoenix Project, that's a bold claim. Here's what caught my attention though—his conference badge didn't say "Author" or "CTO." It said "Author & Vibe Coder." Gene Kim is embracing this new paradigm himself. Meanwhile, I'm watching talented engineers resist changing how they've always worked. Yes, there are rough edges. Yes, things need to be ironed out. But that's exactly what people said about DevOps in 2013. The Phoenix Project showed us that dramatic transformation was possible in how we build and deploy software. Something similar is happening right now—but it's bigger, way bigger. I'm at the AI Engineer Code Summit this week—a gathering of the leading minds and companies pushing the boundaries of AI in software development. The main sessions start today. The Phoenix Project wasn't about removing friction from deployment. It was about fundamentally rethinking how we deliver software. DevOps required a new way of thinking. This next wave is the same. It's not about making the SDLC faster—it's about transforming it entirely. Planning. Architecture. Code reviews. Testing. Deployment. The whole process needs to be reconceived. And here's what most people aren't saying out loud yet: eventually, AI won't just be assisting. It'll be in the driver's seat. That's what AI Native means to me. The question isn't whether this changes software development. The question is: will you be learning it, or learning about it from your competitors?

The more essential you make yourself, the more you've failed as a leader. That's a lesson that took me years to embrace. Fifteen years ago, I was teaching Marco Salazar how to estimate software projects. Today, I trust him with decisions I wouldn't have delegated to anyone back then. His work running software delivery has been invaluable. The same pattern repeated across Devsu: Santiago Burbano and I built our recruitment process together years ago. Now he owns it completely - I haven't touched recruitment in ages. Marco Antonio Balboa transformed how we think about IT and cybersecurity. Ana María Lacouture M. built our people operations from the ground up. Anthony Suárez is leading our AI hub and pushing the boundaries of what's possible with AI in Velx. And these are just a few examples - there are so many others across the organization who have taken ownership and grown beyond what I could have done alone. Each of them started somewhere specific. Each grew into something more. Most founders fear this. We think our value comes from being needed. From having the answers. From being the bottleneck everyone waits on. But there's an ancient story that reframes this completely. Moses was burning out, trying to judge every dispute himself. His father-in-law Jethro watched and said: "This isn't sustainable. You need to teach others and delegate, or you'll wear yourself out." Moses listened. He built a system where capable people handled most decisions. He became unnecessary for day-to-day operations. And that's when the real work began. Your job as a leader isn't to be indispensable. It's to build people and systems that don't need you. To create something that survives and thrives beyond your involvement. The paradox? The less they need you, the more you've actually accomplished. What's been your experience with delegation and building independence in your team?

I used to love recording training sessions for new hires at Devsu. Now I realize traditional training can't keep up with what AI makes possible. At Devsu we built Velx to modernize legacy systems faster with very strong results in our initial implementation group. But here's what I noticed: Properly prompted, the AI processes knowledge quickly, but no engineer can keep up with that speed. Engineers still need to validate and understand outputs before working with them effectively. The bottleneck moved. It's not processing power. It's how fast humans can learn. AI eliminates repetitive work, yes. But it creates a new challenge: humans learning what's now possible can't keep pace with what AI can already do. These days, my training happens differently—pairing with people, checking their work, providing feedback. I wish I had more time for it, because leveling up talent is what unlocks everything. Once we figure out what and how, we can train a model in hours, but we can't train humans in hours. That gap? It's only getting wider. "Wise people listen and increase their learning." — Proverbs 1:5 (NLT) The companies that win won't just have the best AI. They'll have people who can learn fast enough to use it. I'd love to discuss this topic. What are you seeing in your teams?

AI can analyze a massive codebase in minutes. It takes your team months to truly understand what they're modernizing. That's the real bottleneck. We're not limited by technology anymore. We're limited by how fast humans can learn. At Devsu, while building Velx, we kept hitting the same wall: AI agents would either drift off-course with too little context, or get overwhelmed (and so would our engineers) with too much information dumped at once. The breakthrough? Stop treating context like a data problem. Start treating it like a learning problem. Context Engineering as a learning problem? Yes. Traditional approaches fail because they assume more information = better results. But both extremes break down: → Too little context: AI agents deviate from business and technical intent → Too much context: LLM performance drops AND humans can't grasp it all Our solution: Progressive disclosure. Instead of overwhelming users and agents with everything upfront, Velx collaborates with your engineers to build understanding layer by layer. Each level gets understood and approved before moving deeper. We're constructing knowledge together, not dumping documentation. The result? We've reduced modernization implementation time by 70%. Not because the AI got smarter. Because we matched how humans actually learn. The models are advancing exponentially. Human learning capacity isn't. That gap is where the next wave of innovation lives—not just in modernization, but across every domain where AI meets human decision-making. This is why I'm so focused on this problem now. I'd love to hear your thoughts: How are you handling the learning curve as AI capabilities outpace human adaptation? (And if this resonates with your modernization challenges, let's talk.)

I've never seen a successful startup founder who waits for permission. Yet I see hundreds who wait for "the right time." Two moments I'd like to share: 2010 – The world was still recovering from crisis. Banks were fragile. Companies freezing budgets. Everyone said: "Wait for stability." We saw it differently. Economic chaos meant opportunity. Companies had reduced budgets but still needed talent and didn't know how to hire reliable tech talent abroad. So Felipe Cornejo and myself founded Devsu in the middle of the storm. Fast forward to 2024 – Staff augmentation stopped growing at the same rate after the exponential growth we saw during the pandemic. This, along with the AI revolution, made me think if it was time to create something new. My team was buried in application modernization projects. We could've kept doing it the "proven way" and stayed comfortable. Instead, we built Velx. An AI-powered solution that cuts modernization time by up to 70%. I had to convince my co-founder and leadership team. Not because the idea was bad—but because everyone was too busy working to lift their heads and learn a more effective way. Here's what I learned from an ancient text—Proverbs 6:6-8 (NLT): "Take a lesson from the ants... Learn from their ways and become wise! Though they have no prince or governor or ruler to make them work, they labor hard all summer, gathering food for the winter." Nature's blueprint for entrepreneurship. No oversight needed. Just relentless proactivity. Stephen Covey nailed this in The 7 Habits of Highly Effective People—proactive people shape their circumstances instead of being shaped by them. They're not victims of their environment. They're architects of their future. This isn't just about business. It's about life. Wherever you are—at home, at work, in your community—you can choose to be proactive and create a better reality. You can wait for circumstances to change, or you can act now with what you have. The photo is in 2011, when I was visiting one the first Devsu clients, Minh L., along with a good friend Sebastian Contreras 🙂 Your turn: What's the one thing you've been waiting to start? Share it in the comments—I'd love to hear your ideas and provide feedback. I'll personally reply to each one.