Speechify Boosts Speed with Vercel and Next.js

Those 500,000+ pages to 60 million users were also served across 40+ languages and they cut costs 50% by auto-scaling with Fluid compute. If you want to start building on ▲Vercel, reach out.

You've used Vercel for years. Someone asks what Vercel Agent does.   You've never heard of it.   Vercel has 27 products. You've probably touched three - Next.js hosting, preview deploys, serverless functions. Fair, that's what most of us signed up for.   But 9 of those 27, a third of their lineup, are built for AI agents. Chances are you've barely touched any of them.   A sandboxed VM for agent code. Short-lived credentials instead of a permanent key in an env var.   An agent that investigates your production alerts and waits for your approval before touching anything. Swipe through the map of all 27, then the deep dive into the 4 most likely to come up in your next system design conversation.   Save it for the next "what does Vercel even do" conversation.   Practice real front-end system design questions : https://lnkd.in/dVPu2yPH

🔧 A while back I wrote about how a broken WebSocket in a Google Meet add-on led me to Google's A2UI protocol, and later about the shift that mattered more: substrate, not composites — trading hardcoded UI blocks for a catalogue of atoms an agent composes from. 📈 That catalogue kept growing — and it now has a proper home: a2uicatalog.ai. 468 atoms with a live preview on every card, a working MCP endpoint, and enough problems solved (and not solved) that it needed a proper write-up rather than another thread here. 📝 First post is live: "Building in the Age of Frontier Models: Why Frugality is My Go-To AI Architecture" — five problems nobody asked me to solve, and the idea running through all of them: spend the expensive, non-deterministic frontier model once, and make it produce something that replays for free, forever. I've started calling it artifact distillation. 🔗 https://lnkd.in/egWGnvhE 🤝 MIT-licensed, contributors welcome. #A2UI #MCP #GoogleWorkspace #MCPAPPS

Do people not notice this? NVIDIA is giving away free API access to DeepSeek V4 Pro, GLM 5.2, and the  550B Nemotron 3 Ultra — on their own GPUs. No credit card. Two minutes. Yeah seriously, its free.Every free tier in AI has died. Google's. OpenAI's. This one is still live. Now. Don't waste it on a chat box. Kondi (free, open source) turns that one key into a full AI team: Nemotron plans, GLM codes, DeepSeek reviews. Use it or dont but woah. https://build.nvidia.com https://lnkd.in/eAYzGG-4

If you're running production apps on Grok 4.5, you've likely felt this: API caps just got cut in half across the board, and Pro tier subscriptions lost their fixed query limits entirely in favor of a credit-based sandbox. At $8.80/million tokens, that's not a small adjustment it's a budget emergency for teams that haven't changed how they manage context. The core issue: most teams are burning premium credits on tokens that don't need premium reasoning. Full conversation histories, entire log files, and boilerplate are all getting sent to an expensive model when a fraction of it actually matters. Here's the orchestration approach that's reportedly cutting overhead by up to 60%: 📉 Sliding context windows cap session length (e.g. 30K tokens), then auto-compress older turns into a summary instead of replaying full history every request. 🧹 Semantic redundancy filtering strip empty lines, boilerplate, and irrelevant log noise before anything hits the API. Reasoning models need the signal, not the noise. 💾 Prompt caching store static system instructions and schemas server-side so you're not repaying for the same context on every call. 🔀 Multi-model fallback routing when you hit a rate limit or error, automatically reroute to a cheaper capable model (GPT-5.6 Sol Max or Qwen 3.7 Max, depending on budget) instead of letting the app go down. The cost delta is stark: a naive pipeline averages ~$880/month at 10M input tokens. An optimized hybrid setup with routing brings that to ~$340. Full caching gets it down to ~$120. The bigger lesson applies beyond Grok: treating any frontier model as a general-purpose conversational engine is an expensive habit. Active context management isn't optional anymore it's the difference between a sustainable AI budget and a runaway one. Full breakdown: https://lnkd.in/d_jevpPM #AI #LLM #APICosts #TokenOptimization #AIInfrastructure

The internet is quietly splitting into two. Last week, Cloudflare introduced Kitesurf, a browser built specifically for AI agents instead of humans. At first glance, it looks like another browser launch. But it signals something much bigger. For decades, browsers have been designed around human behaviour. Humans need tabs, smooth scrolling, animations, videos, accessibility features and pixel-perfect rendering. We browse, read, compare and decide. AI agents browse differently. They don't care about beautiful interfaces or animations. They need structured data. They need to execute JavaScript, understand the DOM, click buttons, fill forms, extract information and move on. They need speed, isolation and efficiency over visual experience. The same web. Two fundamentally different users. Cloudflare's launch feels like one of the first public acknowledgements from a major infrastructure company that building for humans and building for agents are becoming two separate engineering problems. Which raises some interesting questions. → If browsers are evolving for AI agents, what happens to the rest of the stack? → Will APIs become more agent-first than human-first? → Will products need separate experiences for people and autonomous agents? → Will every SaaS product eventually need to answer not just "Is this easy for users?" but also "Can an AI agent reliably use this?" And perhaps the biggest question for founders building today: When you're designing your next product, who are you designing it for? Only humans? Or humans and AI agents working together? The answers to those questions may shape the next generation of digital products more than the next AI model release. PS: On a completely unrelated note, here’s a picture of our two builders collaborating to fix a bug on a demo environment. Question to you? Is this image for humans or for bots? #SP18 #SP18Builders #AI #AgenticAI #ProductDevelopment #SoftwareEngineering #CloudComputing

Exciting day here at Cloudinary. We just became the first visual media platform purpose-built for agent-native development. Built on infrastructure Cloudinary has been developing since 2012, AI agents can now find the right asset, govern how it's used, and act on it, without needing a human in the loop at each step. For developers, it means less time wrestling with the mechanics of media integration and more time on application design and product decisions. For anyone building AI-powered applications, learn more and let us know what you think! https://okt.to/mHXUL2

The most valuable thing a senior dev does with AI tools? Say no. Every AI coding assistant has the same blind spot: it wants to handle every case. Ask it to write a function, and it'll add input validation, retry logic, and a fallback for a scenario that will never happen in your codebase. Ask it to fix a bug, and it'll refactor three files "while it's in there." None of this is malicious. The model is pattern-matching on what "robust code" looks like in general — not on what your app actually needs. This is where the dev's judgment matters more than the dev's typing speed. A few things I've started doing: → Ask "will this case ever actually occur?" before accepting a new fail condition. Most obscure edge cases the model surfaces are theoretical, not observed. → Push back with "keep it simple" or "don't add error handling here" as explicit constraints — not just accept the first draft and clean up later. → Treat every suggested abstraction as a cost, not a free upgrade. Code that handles 10 cases you don't have is harder to read than code that handles the 2 you do. → Remember that "more defensive" isn't the same as "more correct." A codebase where every function guards against inputs that can't occur is just noise wearing the costume of rigor. The models are excellent at generating code. They're not yet good at knowing when to stop. That's still on us. Complexity is easy to add and expensive to remove. The best use of a senior dev's time with AI tools might just be saying "no, we don't need that" — on repeat. https://lnkd.in/gcq63gTt #AIEngineering #SoftwareEngineering #CodeQuality

Vercel just published a deep dive on the website I work on every day, and it's been exciting to see some of the engineering behind it shared publicly. Over the past couple of years, I've had the opportunity to help build the Speechify website—a platform that's grown far beyond what most people think of as a marketing site. Today, it serves hundreds of thousands of dynamic pages across dozens of languages, supports an experimentation-heavy growth engine, and reaches millions of users worldwide. Some of the engineering problems we've solved include: - Scaling to 500,000+ pages across 40+ languages. - Serving highly dynamic content while keeping pages fast. - Building fully editable content and page layouts directly from Payload CMS. - Running continuous product experiments without sacrificing performance. - Leveraging Next.js caching, ISR, and Cache Components to keep infrastructure efficient at scale. One of the most rewarding parts of working on the platform has been building reusable systems instead of one-off pages. We've invested heavily in a component-driven architecture where content and layouts are fully editable through Payload CMS, enabling teams to create and iterate quickly while keeping the codebase consistent, scalable, and performant. The platform is also deeply experimentation-driven—we rely on it to continuously test ideas, validate hypotheses, and ship improvements quickly, all while maintaining a fast and reliable experience for millions of users. Huge kudos to everyone across Engineering, Design, Product, SEO, and Content. Building a platform at this scale is a true team effort, and I'm grateful to have worked alongside such an incredible group of people. If you're interested in how we built and scaled a modern Next.js platform for millions of users, the Vercel case study is well worth the read. https://lnkd.in/dZR74TbE - Written with Speechify Voice Typing

Most enterprise AI is trapped in the "chatbox trap," forcing users to watch slow, token-chugging text streams just to check a database record. It is slow, inflates your LLM bills, and ruins the user experience. The way out? MCP Apps (Server-Rendered Agentic UIs) using the Model Context Protocol. We just published a full technical breakdown of this architecture, covering: The stateless spec updates that solve server memory bloat. Context Poisoning security risks in your database. The React component sync pipeline required to prevent UI crashes. Stop overpaying for tokens to deliver a slow product. 👇 Read the full technical breakdown here: [Link in first comment] #ModelContextProtocol #MCP #GenerativeAI #ReactJS #CTO